A live "vibe check" of OpenAI's new GPT-5.5 model by the Every team, comparing it against Opus 4.7 on speed, usability, and performance across coding, dashboards, writing, and enterprise workflows. # Channel Identity Observations: Every ## Explicit Self-Description - **Subscription positioning**: The channel repeatedly states a tagline: > "Every is the only subscription you need to stay at the edge of AI." This is stated at least four times throughout the stream. - **Scope claim — ideas, apps, training**: The host describes the offering as: > "We do ideas, apps, and training... Ideas, apps, and training all part of one subscription." Including a daily newsletter ("you just read one thing and you are absolutely caught up"), a suite of bundled apps, live streams, camps, and courses. - **Early-access claim**: The channel claims privileged pre-release testing access: > "We get these models way before they're released and on release day you can expect a vibe check like this from us." - **Product suite named on-stream**: Cora (AI email agent), Monologue (dictation/notes), Sparkle (file organizer), Spiral (AI ghostwriter), Proof (markdown editor), PlusOnes (hosted OpenClaw) — all described as included with subscription. - **Consulting/enterprise offering**: Mike Taylor describes the practice: > "A company will come to us and say, 'We just bought Claude Code for this whole team. How do we get them to use it?'" With "hands-on tactical" workshops and an exploration of subscription-based consulting pricing. ## Claimed vs. Observable Scope | Claimed | Demonstrated | |---|---| | "Stay at the edge of AI" | Team members describe 3 weeks of pre-release testing of GPT-5.5 (host: "We've been testing it for the last 3 weeks or so") | | Coding + writing + design + marketing coverage | Panel includes engineers (Naveen, Kieran), a writer (Katie Parrott), head of growth (Austin), and consultant (Mike Taylor), each presenting role-specific tests | | Training/consulting for enterprises | Mike references live client work: "we're working on AI adoption with a few companies trying to get the teams unblocked" | ## Intended Audience - **AI practitioners across skill levels**: Austin explicitly frames the non-engineer persona: > "For someone like me without like a technical or engineering background I feel way more comfortable doing any kind of engineering work with the latest OpenAI models." - **Knowledge workers and builders**: Use cases span engineering, writing, growth/marketing, OKR management, and design. - **Paying subscribers as the commercial target**: Repeated calls to action ("you should go to every.to and you should subscribe"), with free and paid tiers of the vibe check noted ("A lot of it's free. Some of it's for paid subscribers"). ## Observable Presentation Patterns - **Recurring format — "vibe check" on model release days**: The host frames this as a repeatable series: > "We do these vibe checks every time a new model drops." Terminology like the "reach test" ("do you reach for it every day?") and green/yellow ratings suggest an established, codified methodology. - **Proprietary benchmarks demonstrated on-stream**: A "senior engineer benchmark" (GPT-5.5: 62 vs. Opus 4.7: ~33), an underspecified "Talkform" vibe-coding benchmark, a persona-cloning benchmark, and dashboard/PowerPoint design tests. - **Team-based, multi-perspective reviews**: Disagreement is showcased rather than smoothed over — the host is "green," Kieran and Mike are "yellow" on the model, and Kieran's generalist critique is aired at length. - **Guest access to industry insiders**: OpenAI's Romain Huet (head of developer experience) and Dom (Codex DX) appear to discuss the model and demo capabilities. - **Live commerce integration**: The stream functions simultaneously as editorial review and subscription funnel, with extended promotional segments for Every's apps and consulting between demo segments. ## Evidence Notes - No curated metadata was supplied; all observations derive from title and transcript. - Team roles are stated on-air by the host or guests themselves (e.g., "Mike Taylor, who's our head of AI consulting"; "Kieran, GM of Cora at Every and the creator of Compound Engineering"; "Katie Parrott, staff writer at Every"). # Delivery Observation: LIVE VIBE CHECK — GPT-5.5 ## Format & Structure - **Live-stream panel format** with a rotating cast of co-hosts and guests; the host (Dan) orchestrates handoffs by name: *"Mike Mike Taylor, come on. Come on over."* - **Recurring segments**: benchmark recaps, guest demos, product plugs for Every, and a running "reach test" framing — *"which is do you reach for it every day?"* - **Color-coded verdict system** used as a summarizing device: *"I'm a green... Kieran though is a yellow."* ## Tone & Register - Enthusiastic, casual, profanity-adjacent delivery (bleeped in transcript): *"And it's really [ __ ] great."*, *"I [ __ ] love model release days, honestly."* - Conversational hype words recur: **"beast," "workhorse," "daily driver," "vibe coded."** — *"this model is a beast"*, *"it has become my daily driver."* ## Recurring Phrases - **"Here we go. Here we go."** and **"Okay, here we go"** — repeated verbal reset markers at transitions. - **"It is out. It is out."** — immediate repetition for emphasis; this doubled-phrase pattern is frequent: *"It's really great"* / *"It's really great."* - **"I love it."** — appears at least four times as a segment-closing reaction. - Repeated verbatim marketing tagline: *"Every is the only subscription you need to stay at the edge of AI."* (used at least three times). ## Teaching Habits - **Quantifies with benchmark scores repeatedly**: *"GPT-5.5 scored a 62 out of 100. By comparison, GPT Opus 4.7 scored a 33."* — restated nearly identically at least three times. - **Uses analogy to explain model differences** (guest habit): Mike's *"I felt like comfortable and safe like getting into a Waymo... Opus is the Tesla, you know? It's like a little dangerous."* - **Invites demonstration**: *"Can you show us some of the apps"*; prompts guests to share screens and narrate. - **Flags caveats explicitly**: *"There's an interesting caveat though. Do not throw out Opus 47 yet."* ## Communication Patterns - Heavy use of filler words ("um," "uh," "like," "you know") across all speakers; false starts are common: *"It's a It's a"*. - **Humor and banter** as connective tissue: *"you had pink eye so you were... kind of out of commission"* — *"Pink eye really helped me vibe code"* [laughter]; the closing "stay hydrated... keep your agent liquid cooled" bit. - **Direct handoff questions** to guests: *"Kieran, what else do you think you've experienced that we haven't talked about yet?"* - Occasional self-aware hedging on anthropomorphizing models: *"I hate to anthropomorphize, but that's how it feels."* (Mike, twice) ## Stable vs. Isolated - **Stable patterns** (high repetition): doubled phrases for emphasis, benchmark score recitation, the Every tagline, guest handoffs by name, "I love it" reactions. - **Isolated moments**: the hydration/Coke Zero joke, the "Romain is not ready" guest-timing stumble, Katie's *"I'm not ready at all"* — one-off, not indicative of a pattern. - **Uncertainty note**: speaker attribution between the host and guests is occasionally ambiguous in the transcript (overlapping talk), so a few quoted lines are attributed by context rather than explicit labels. # Observation Record: LIVE VIBE CHECK — GPT-5.5 *Positions and claims as expressed by speakers in the transcript; no inference beyond stated content.* ## Participants - **Dan** (host, Every) — ran the stream and the senior engineer benchmark - **Mike Taylor** — head of AI/technology consulting at Every - **Kieran Klassen** — GM of Cora at Every, creator of Compound Engineering - **Naveen** — GM of Monologue - **Austin** — head of growth at Every - **Katie Parrott** — staff writer at Every, author of the vibe check - **Romain Huet** — head of developer experience at OpenAI - **Dominic ("Dom")** — leads DX efforts for Codex at OpenAI ## Release Status & Availability - **Dan:** GPT-5.5 is out live in Codex and ChatGPT; not yet on the API. He speculates rollout over hours and says OpenAI is holding the API back for testing because it's "a very powerful model," which he calls "probably good" and "seems to be the new standard." - **Romain:** OpenAI is taking a "very safe approach" and expects API access "extremely soon, hopefully in days, maximum," pending safeguards. - **Dan:** The model is a totally new pre-trained model — "not a fine-tune of a previous GPT-5 version" — the "Spud model." ## Benchmark Findings ### Senior Engineer Benchmark - **Dan:** GPT-5.5 scored **62/100** (best score) vs. **Opus 4.7 at ~33** — nearly a 30-point swing. The benchmark tests rewriting a vibe-coded "slop codebase" the way a senior engineer would, compared against rewrites by two real senior engineers. - **Key caveat (Dan):** 5.5's best score came **using a plan written by Opus 4.7**. "Do not throw out Opus 4.7 yet." Opus's plans are still better; 5.5 is the model you want executing/coding. ### Reasoning about the difference - **Dan:** Given an underspecified "rewrite from first principles" prompt, both models can produce a plan, but Opus 4.7 picks off only a small piece and patches rather than rewriting; GPT-5.5 executes the full plan over many turns/tokens, with the "courage or assertiveness" to delete code. This behavior appears **only on extra-high reasoning**, and works best with Opus's terse, spec-like, "contract-heavy" plans. - **Mike:** Opus feels like it "panics" on long conversations ("do you want to wrap this up?") — he notes he's anthropomorphizing. Codex/5.5 is "dogged and determined." ## Reach Test (Daily Driver Ratings) - **Dan:** Green — daily driver for everything; still occasionally uses Claude on mobile. - **Naveen:** Green — used ~900 million tokens; didn't reach for other models, including for support replies and front-end work he previously sent to Claude. - **Katie Parrott:** Green (per Dan). - **Kieran:** Yellow — daily driver remains Opus 4.7. - **Mike:** Yellow — still uses Opus when micromanaging tasks. ## Individual Positions ### Mike Taylor - Calls 5.5 "the most reliable model I've tested"; analogy: 5.5 is like a **Waymo** (safe, delegable), Opus is a **Tesla** (a little dangerous). - Used it to build training curricula from call notes; "hasn't failed on that once." Opus produced "cool" titles, but for corporate training you want dependable and unoffensive — 5.5 wins there; Opus for marketing copy. - **Persona cloning benchmark** (from his failed startup cloning focus-group participants): 5.5 initially refused to impersonate him ("alignment checks") but was quickly fixed and worked "beautifully." Opus "tried too hard," over-emphasizing his verbal ticks; 5.5 was "much cleaner." - **Writing:** 5.5 is dependable and won't say something stupid, but he still prefers Opus's wit/neuroticism; Opus 4.7's flaw is saying something clever then stopping early. Calls 5.5 a true "workhorse." - **Dashboards:** 5.5 beat Opus on a custom dashboard (more professional, tasteful color scheme) — he wasn't expecting this since Opus made the best PowerPoint deck. Contradiction he flags: 5.5 failed the PowerPoint test but won the dashboard — "it's not a clear picture" on design. - On OpenAI's computer use: he wants to try it; it's also what enterprise clients are "most scared of." ### Kieran Klassen (Yellow) - In his benchmarks, 5.5 scored the same as 4.7 on coding — "a very good coding model." - **Core critique:** 5.5 "feels more like a specialist and less like a generalist." Claude/Opus is the generalist — good at product work, details, and big picture. 5.5 is strong at execution and details but "sometimes breaks down if you look at it from far away... things not being coherent." - Frames this as philosophy, not quality: OpenAI and Anthropic have different views of engineering work; different people align with each. Naveen is an "engineer engineer," Kieran a "product engineer generalist." - **Design:** typography and structure improved; visual design degraded (right side of a test layout "doesn't make any sense"). Notes it previously added unnecessary UI elements — that's gone. Opus 4.7 did better on an "island test" (colors, animations). Also notes his benchmark tests the harness (Codex CLI, plugin framework, compound engineering), not just the model. - **Ruby:** 5.5 is "just not good at Ruby" — "it does TypeScript stuff but then in Ruby," which is his "biggest blocker." - **Successes:** one-shot an entire rubber-ducky store web app (first time ever with 4.7/5.5 class); one-shot a React library ("riff-react") over three commits with useful proactive suggestions (e.g., consent screen). Notes it took ~3x more tokens in planning/review states. - **On his yellow:** says yellow also reflects the harness; recalls rating Cursor yellow before using it daily. "I'm very enthusiastic about this model" and is running three Codex tasks in the background. Says 5.3 was when OpenAI "got my attention." - After the OpenAI segment: his own image-to-website benchmark performed badly (5.5 just used the image as a background, on extra-high reasoning) — concludes results depend on harness and prompting, wants to learn to get more out of it. ### Naveen (Green) - Previously agreed with Kieran about older GPT models, but 5.5 is "really well rounded." He now uses it for everything including web/front-end and support replies he used to route to Claude. - **Vibe coding:** built three apps, including "Dayline" (a Raycast-alternative to-do Mac app) entirely from a screenshot prompt, in a **single ~200-million-token thread**, with no code review; also produced a synced iOS app. Doesn't plan to release it. - Added remote MCP support to Monologue by asking 5.5 to plan across front end, back end, iOS, macOS, then saying "Implement this plan" — it worked continuously across multiple codebases. Credits good context handling across compaction. - **Design following:** counterexample to Kieran — implemented Daniel's design for Monologue nearly one-shot with a Playwright-screenshot feedback loop in about an hour. ### Austin (Head of Growth) - Switched from Claude Code to the **Codex app** as daily driver, largely due to app speed vs. the "clunkiness and slowness" of the Claude desktop app. - Previously felt codex "makes me feel like I'm dumb" — planning mode asked questions with unexplained options and a dismissive tone. "All of that is gone" with the new model; he now trusts and understands it. - Example: pointed the model at Notion and Slack, brain-dumped campaign goals for the Plus One launch; the plan going to market is ~90% the model's output, informed by team taste in meeting notes. - Values the app recommending and building automations he uses daily. - Would only switch back if the Claude desktop app got dramatically faster or Opus far surpassed OpenAI models for knowledge work — currently they "feel about even for most stuff besides design," where he still searches for Claude for video clipping/image generation. - Says speed enabled doing knowledge work while watching NBA playoffs at "10% brain." - **Contradiction on OpenAI media:** with identical prompts, OpenAI models just "zoom in" on screen recordings for social assets — "the laziest possible approach" — while Opus engages creatively. But the new GPT image generation tool for YouTube thumbnails was "so good": near-perfect logo, correct finger count, strong one-shot thumbnails. ### Dan (Host) — Additional Positions - Headline: "GPT-5.5 is OpenAI's workhorse model." - Rare for a model to be both strong at senior engineering and a good, personable, fast collaborator — improving on both dimensions at once is unusual (contrasts with Opus 4.7, which he says got better at programming but "lost some of its empathy"). - "Speed is in some ways a type of intelligence." - Notes front-end image-model-then-implement workflow "short circuits" 5.5's weaker taste. ## OpenAI Team Claims ### Romain Huet - Internally at OpenAI, adoption of the Codex app has spread beyond engineers to adjacent roles; model is "smarter at any kind of professional task," and the combo with Codex's tools/plugins (~100 plugins, artifacts, Slack/Notion integration) makes it a "productivity app for much more than just a developer." - Mentions Chronicle (research preview, watches your screen) and computer use — claims no other implementation lets you keep using your computer while Codex runs tasks in the background driving native Mac apps. - Relayed feedback: developers report not specifying tasks much anymore and trusting the model to navigate large codebases; an engineer at Ramp (Will) reported 5.5 was "complete plug and play" in their custom harness and discovered new tools (database access) other models couldn't find. - On the image-model-to-frontend workflow: acknowledges front-end/taste "is still something we have to improve" and they're "pushing on that very hard," but ImageGen 2.0 + 5.5's thoroughness lets you "iterate on your design first... then get it done with GPT 5.5." ### Dominic - Built ~50 demo apps in a week, including one-shotting "a full 3D modeling software" at 11 p.m.; it already included 3D file download support unprompted. - **Artemis 2 demo:** from a screenshot plus WebGL instruction, 5.5 fetched real NASA flight data, cited sources, rendered the trajectory, and even corrected illustrative (non-real) planet scaling from the source image; added a true-scale mode on request. - **Front end app builder skill:** silently launched; combines ImageGen design with 5.5 implementation; one-shot a GPT-5.5 launch page from the blog post. The model used in-app browser use to take screenshots, check mobile/desktop, and self-verify "until it's satisfied." - Day-to-day: improved Codex docs by having 5.5 pull a component from a different codebase and draft examples from a Google Doc; typical prompt is now "go look at the last feedback Romain sent me on Slack and go fix that and send him a message when the PR is up." ## Katie Parrott (Writing) - Hadn't used OpenAI models for writing in nearly a year (since ~Sonnet 3.5); canceled her personal ChatGPT subscription ~3 months ago. Initially unimpressed with 5.5 until she "got her hands dirty" and saw how it took direction. - Found the writing **smoother**, logical flow more natural, **less clever** (a plus for dependability). On request it captured her "neurotic energy." - **Speed and intuitive direction-taking:** 5.5 picks up meaning from poorly phrased, minimal feedback, whereas Opus 4.7 "needs a lot of specificity." - Says she is "thinking about breaking up with Claude" for personal use, but keeps both subscriptions through work. - Wrote the vibe check itself in Codex with 5.5; praises the Slack integration for pulling context during a last-minute revision. Also built an OKR-managing agent connected to Notion and calendar. ## Preserved Tensions & Qualifications - **Opus 4.7 is not obsolete:** best 5.5 benchmark result required an Opus-written plan; Opus still preferred for planning, wit, visual design, video/image creative tasks (per Dan, Mike, Kieran, Austin). - **Design picture is mixed:** Kieran and Austin report Opus better at visuals; Mike's dashboard and Naveen's design-following results favor 5.5; Kieran's image-to-website benchmark contradicted Dom's demo — Kieran attributes this to harness/prompting, not the model alone. - **Reach test split:** greens (Dan, Naveen, Katie) vs. yellows (Kieran, Mike), with Kieran explicitly saying "it's not good or bad, just different" work styles. - **Senior engineer benchmark scores reported inconsistently by Dan** across the stream: 62 vs. "a 33" once, "a 30, 32, 33-ish" later — recorded as expressed. - **Anthropomorphizing flagged as such** by speakers (Mike on Opus "panicking"; Dan on 5.5's "courage"). # Speaker Attribution & Provenance Record **Video:** LIVE VIBE CHECK: GPT-5.5 Has it all **Video ID:** 01a0f15a-88c8-7c46-a4e7-393bd4cb3c36 **Format:** Live stream ("vibe check" day-of-release review of GPT-5.5) ## Speakers & Roles ### Host / Creator - **Dan** *(surname not given in transcript)* — Host, creator of the stream; identifies as running Every's vibe checks. Presents the senior engineer benchmark results and moderates all segments. ### Every Team Members (Creator's Colleagues) - **Mike Taylor** — Head of AI/technology consulting at Every. Shares knowledge-work testing (dashboards, persona-clone writing benchmark), and later describes Every's consulting offering. - **Kieran Classen** — GM of Every's Cora product; creator of "Compound Engineering." Rated **yellow** on the reach test; demonstrates design benchmarks (LFG bench, rubber ducky store app, island test, riff-rec React library). - **Naveen** *(surname not given)* — GM of Monologue. Rated **green**; demos the "Dayline" vibe-coded Mac app (single ~200M-token thread); reported ~900 million tokens used on GPT-5.5 over recent weeks. - **Austin** — Head of Growth at Every. Describes switching from Claude Code/Claude desktop to the Codex app; demos YouTube thumbnails made with the new GPT image tool. - **Katie Parrott** — Staff/lead staff writer at Every; **author of the published vibe check**. Rated **green**; discusses writing testing; describes writing the vibe check itself with Codex. - **Laura** — Staff writer at Every; briefly brought on live near the end; no substantive contribution recorded. ### Third-Party Guests - **Romain Huet** — Head of Developer Experience at OpenAI. Discusses model capabilities, Codex combo, API timing. - **Dominic (Dom) Kunda** — Leads DX efforts for Codex at OpenAI. Demos Artemis 2 WebGL visualization and the "front end app builder" skill using ImageGen. ## Promotional Segments (Flagged) - **Every subscription promotion** — Repeated verbatim tagline: *"Every is the only subscription you need to stay at the edge of AI."* Appears at least three times, including a dedicated mid-stream segment touring Every's products. - **Product bundle promo** — Extended tour of Every apps: **Cora, Monologue, Sparkle, Spiral, Proof (proofeditor.ai), PlusOnes**; link: `every.to` / `every.to/subscribe`. - **Vibe check promotion** — Repeated calls to read the published vibe check at `every.to/p/gpt-5-5` (part free, part paid). - **Consulting promotion** — Mike Taylor describes Every's enterprise AI transformation/adoption consulting and a subscription-based consulting model in discussion with "four or five companies." - **Sponsor-like bits** — "Stay hydrated / keep your agent liquid cooled" recurring gag (no named sponsor identified). ## Referenced Sources & Claims ### Internal (Every) Sources - **Senior engineer benchmark** — Every's own benchmark; GPT-5.5 scored **62/100** vs. Opus 4.7's **~33/100**; caveat: 5.5 scored best when executing a plan written by Opus 4.7. - **Reach test** — Daily-driver usage ratings: Dan (green), Naveen (green), Katie Parrott (green), Mike (yellow), Kieran (yellow). - **Kieran's LFG bench / compound engineering harness** — tested Codex CLI harness, plugin framework. - **Mike's virtual focus group startup persona clones** — personal prior project; clone-of-self writing benchmark. ### Third-Party Claims (via OpenAI guests) - OpenAI blog post announcing GPT-5.5 (referenced, "on their site"). - GPT-5.5 is a **new pre-trained model** (not a fine-tune), identified as the "Spud" model (host's claim). - **API availability:** not yet on API at time of stream; Romain Huet says expected "hopefully in days, maximum," pending safety rollout. - Concurrent launches referenced: **Imagine 2.0 / image generation tool, Chronicle (research preview), computer use, Codex app artifacts, ~100 plugins, ChatGPT agents**. - Third-party anecdote: **Will, engineer at Ramp** — GPT-5.5 "plug and play" from 5.4 in their custom harness, discovered new tools (Romain's report). - **NASA Artemis 2 flight data** — downloaded by the model for Dom's demo. ## Availability & Uncertainty Notes - Dan's surname is never stated in the transcript. - Opus 4.7 benchmark score stated variously as "33," "32," "30 32 33-ish" — explicit uncertainty by host. - No curated metadata was supplied; comment availability **not stated**. - Some benchmark results attributed to "Mike" vs. "Kieran" vs. "Dan" interchangeably during discussion; attributions above follow each speaker's own statements. Hello everybody. Welcome to vibe check day. We've got a new model coming out. It's actually It is out. GPT-5.5 is out. We've been testing it for the last 3 weeks or so. And it's really \[ \_\_ \] great. It's really great. It's a very exciting model. Um I I've loved it. It has become my daily driver. Uh so, what we're going to do on this stream for you is we're going to go through our vibe check, which we just published on every.every.to. Every is the only subscription you need to stay at the edge of AI. I'm going to go through all my findings. We're going to have other people from the team who are going to come on. Mike Mike Taylor, come on. Come on over. Um \>\> Hey. This is Mike Taylor. Uh What's your What's your best side? Yeah, this is This is good. This is Yeah, this is my best side. Why don't you take that chair? Thanks. Uh so, we've got We've got me. We've got Mike Taylor, who's our head of AI consulting AI tech consulting at Every. Um Mike, you're not quite in the shot. All right. Great. And uh we are going to go through this model for you. So, what I'm going to do is I'm going to share share my screen. And here we go. Here we go. Uh yes. So, GPT-5.5 is out. It is live right now. They're not releasing it on the API quite yet. Uh but, it should be out in your Codex, in your ChatGPT. Um and I I assume that will start rolling out over the next couple of hours. Um they're they're holding it in the API for just a bit uh because it's just a very powerful model and they're doing a lot of testing on it, which I think is probably good. Um and seems to be the new standard for these models. Like we're We seem to be reaching a level where people are like, "Oh, this is kind of worrying." But, okay, here we go. Here's our vibe check. We just published it on every uh every.to/ p/gbt- 5-5. And our headline is GPT-5.5 is OpenAI's workhorse model. Um What's really it So, it is out. It is out. There's a There's a blog post on their site. You You should be able to see it. Um We What we uh what we found in testing this model is it's actually quite rare for a model to be um really good at both senior engineering type tasks and a really good workhorse. And this model is that. It is super collaborative. It's super fast. It's actually like pretty personable. And it did the best on uh our senior engineer benchmark of any model that we've tested. So, on our senior senior engineer benchmark, which tests how good a model is at um rewriting a an existing codebase in the way that a senior engineer would, benchmarked against real on a real codebase that two real senior engineers rewrote separately, um GPT-5.5 scored a 62 as its best score. Opus 47, 62 out of 100. Opus 47 um uh by comparison scored I think its best score was a 33. So, there's like almost a 30-point swing between Opus 47 and GPT-5.5 on this kind of task. There's an interesting caveat though. Do not throw out Opus 47 yet because our uh best the best performance of this model of GPT-5.5 comes from uh using a Opus 47 plan. So, if you use them together, they get super powerful. Um Codex is or GPT-5.5 is the model you want to be coding. Uh but, you actually Opus 47's plans I think are actually still better than than than 5.5. But, uh when you actually put it on this benchmark, it is like miles better at this kind of task. Um so, we have a full We have a full vibe check here. Uh it's uh it's on every every.to/ uh You should be able to see it. every.to/ p/vibe/p/gbt-5-5. Full vibe check is here. Um published already. Um It's a totally new pre-trained model. So, this isn't a fine-tune of a previous a previous GPT-5 version. It's the Spud model that you've been hearing a lot about. It's super powerful. Um But, there are there's sort of like a mix of reactions here. So, it has absolutely become my favorite, my daily driver model. Um it's I use it for everything from my day my day-to-day work to like real engineering tasks. But, other people on the team have different reactions. Mike Taylor, who's our um head of uh technology consulting at Every, has has has had a slightly different um experience with it. Mike, do you want to um just talk a little bit about what you found in testing this model? Yeah, sure. So, um I was also very impressed. I think it's the most reliable model I tested. Uh I felt like comfortable and safe like getting into a Waymo, you know? It's like uh it like feels pretty good. But, I still think like Opus is the Tesla, you know? It's like a little dangerous. Um and when I'm micromanaging tasks, I I still use Opus as my daily driver. Like if I'm heavily involved. Uh but, I would say for um tasks where I know I'm not going to be able to pay that much attention uh and I just need to get done. I want to make sure it's safe. I can delegate it without you know, with peace of mind. I think like this model has it. Um and give me give me an example of like the kind of task that you might want to uh use this model for that uh and we just got Kieran Classen, Kieran GM of Every of Cora at Every and uh the creator of Compound Engineering. Kieran, welcome. Hello everyone. Um Uh so, Mike, just give us give us an example of a task where you're like, "This is actually a really big difference. I might I really want to try this today using this model." that I you might not realize. Yeah, so one thing I've been using it for this week is uh creating the curriculum for training material. Uh because this is quite an onerous task. Uh like you have uh you know, tons of call notes from different people across the organization. Uh like we're we're working on AI adoption uh with a few companies trying to get uh the teams unblocked so they can use more AI. And uh it just requires a lot of diligent work of like we need to go through all the call notes, make sure that like uh all the different themes and issues that we found in those notes have represented in the curriculum. And uh it just like hasn't failed on that once. Uh like I've been uh and and and also I would say that with with Opus I was using Opus for that previously. And with Opus I always felt like I had to like go back through it line by line. Uh and it it came up with some really sharp stuff and like some of the titles were like cool. I'm like, "Oh, I'd love to teach that." But, but actually like you don't want cool titles all the time when you're doing corporate training. Right? You actually want something dependable, reliable that like you know, regular people will find accessible. So, I would say like if I was writing marketing copy, it would be Opus probably. But, uh if it was uh something where I needed to be uh like uh unoffensive, I needed to be um reliable, I needed to capture all the main notes and not be like a little bit wild, then uh yeah, this is this is the model Great. Um so, we've got a couple more people to to add into this vibe check. So, we've got as I said before, we've got Kieran Classen, the GM of Cora, and the creator of Compound Engineering. And we've got Naveen, the GM of Monologue. Hello, Naveen. How's it going? Hi. Hi. Good. Good. Really excited. So, okay. So, I we I We've got if you look at our vibe if you look at our vibe check, um what we uh this comes from um 3 weeks of internal testing. We always do the reach test, which is do you reach for it every day? For And if you do, what do you reach for it for and what do you reach for it over? So, GPT-5.5 is absolutely my my daily driver model. I'm a green. Kieran though is a yellow. So, Kieran, you have some I think you have some some nuanced thoughts on the the strengths and weaknesses of this model. Can you talk to us about what you found? Yeah, absolutely. So, for for me to use a model and use it daily, uh it has to help me build Cora. I'm building a new version of Cora. Um And it's a lot of product work. It's a lot of coding. It's a lot of testing. Uh it's very wide work. Lots of work. So, front end, back end, everything. And what I noticed is that while 5.5 is a very good model. Like in my benchmarks that I ran, um like when when I ran it in 4.7, I was like, "Oh, this is the best coding model out there." And 5.5 GPT-5.5 scored the same on that specific benchmark. So, it's a very very good coding model. Um But, why why don't I use it every day is that it's also it's also it feels more like a specialist and less like a generalist. That's kind of how I look at it. Like Claude is the generalist that is a very good coding model, but it's also very good at product work. It's good at going into details. It's uh it's going it's good at like looking at the big picture. And for um GPT-5.5, it's very good in execution and going into details but sometimes it like breaks down if you look at it from far away and you just see things not being coherent. And I need that. I need a generalist. I need some something that is working for everything. Although I'm using 5.5 for review for execution. Like I use it in my flow. But it's not the daily driver that I go for first. Like the daily driver is still 4.7 because it's better of a generalist and since I do product work and more general work it's it's yeah it's I think it's still the same philosophy as we ever always seen. Open AI just like have a different perspective on what engineering work is than Anthropic and you see that in the model. It's not good or bad. It's just some people are more aligned with one than the other and I think Naveen for example is more aligned with the Open AI take on what engineering work is than my take. Like I'm more of like a product engineer generalist. And Naveen is maybe more of like a engineer engineer and like it's just for different people. It's not better or worse but for me personally while it's a very strong model and I one shot at like amazing things with 5.5. All the models are very good which is a very good luxury we have. Like there's nothing bad about any of these models. They're amazing. Yet like these little things make you go for one or the other. Um so on that note Naveen Jim of Monologue you're you're usually a big GPT fan. So I'm curious what you think of this release and and sort of how you compare and contrast Kieran's feeling about this model cuz I I'm having this experience where I feel like I I totally see some of the things that Kieran says and I I have a slightly different feeling and I think that you have some different feelings so there's like I think there's a lot of different viewpoints and different types of people doing different types of work that make this model feel like holy \[ \_\_ \] \[ \_\_ \] I've never I I've never seen a model do this versus maybe it's not as as good of a generalist for product work like like Kieran's finding. So Naveen tell us what you found in your testing. So I I used to agree with Kieran with the older models of GPT when he says this. Yeah I'm with you because I also used to reach out to cloud code like you know Opus models. But coming to this 5.5 it just feels really well rounded model. Here in my case I'm like you know work in Python code base or Swift code base like native Mac app. When it coming to web app code like code is not really good. So when you're building front end or anything I always reach out to cloud code or some other model. But with 5.5 I didn't reach out to any other model. Like I know Opus 4.7 released. I tried it couple of times but I just went back to 5.5 because it's really great for like everything. Before I I also do support right for Monologue. I do turn off replies. Obviously we use AI to do that. Before I used to use cloud code because it's like really good at writing. But with 5.5 I just completely moved to moved that as well. So and main drawback with the older models is vibe coding where I know for sure I can't pick code X as my vibe coding tool. If I get come across like any other like you know any idea I just always go to cloud code. But in this case as well I vibe coded three different apps in last couple of Can you show us some of the apps cuz cuz one of the thing one of the cool things is that Naveen used like 900 million tokens on GPT 5.5 over the last couple weeks. So he's really really making those GPs go burr. So Naveen can you show us I know you did an app called Dayline. So I'd love for you to show us an app that you were able to vibe code with GPT 5.5 on the side while you're also shipping a bunch of features for Monologue while you were also you had pink eye so you were pretty you were kind of out of commission for the Pink eye really helped me vibe \[laughter\] code this because I'm just you know resting but I have Monologue dictating into you know code X. Let me share my screen. I'm sharing my screen. So first One second one second. Okay. I'm going to remove myself and then I'm going to add you. Good. Okay now we can see your screen. Can you zoom in a little bit? Okay never mind. Okay let me open my downloads. I just like tried to download recent things. So one Okay vibe coding Dayline. Yeah I want \[clears throat\] to show you guys this thing where it's a Raycast alternative. I personally use Raycast notes like this notes where you can just bring this whenever you want to on top. But the thing is this is like a plain text right? I don't want this kind of plain text so I thought okay let me make the to-do list which is day-to-day. Like I can just have this like to-dos on a daily basis. So you can see I'm already using it. Um and yeah this whole thing I didn't look at single code. It just completely I just gave it a screenshot of Raycast. Like okay I really like Raycast like just go take do do it and then it vibe coded. What's really cool thing about it is though the Mac app that you see here like you can see I like I hit enter you can type I hit enter you can go back. All these like interactions minor interactions that I'm using it's like difficult to implement you know in a Mac app. And it did that all in one single thread. This is one single I think 200 million token thread that you're seeing right now. That's 200 million tokens? Oh my god. Yeah. It's all in one single thread. Here I used build iOS apps plugin. That's the plugin that like you know code X comes with. This is my long prompt. I want to implement today. This is what I was doing \>\> \[laughter\] \>\> when I got pink eye. So and then it implemented the first thing. You can see 49 messages. Like Wow. \>\> It just went at it. It implemented the initial version. I was like started oh can you actually take a screenshots on the Mac app? And then do it for iOS as well. There is a iOS app as well that syncs automatically. I'm just like blown away with like you can see I just keep on having this to and fro and I'm just now regularly using it. I don't think I will be releasing it for anyone else but it's just fun to test to test it out. And today I experienced one bug and I just said that it fixed it and I'm just running it again. That's it. That's really that's really interesting. Yeah I think there there's something really interesting here I want to pull out because because I've done a lot I've done a lot of testing with this and I think some of your results are similar to mine. And the the some of the things that I found are so I have this senior engineer benchmark. And like I said if you test GPT 5.5 on the senior engineer benchmark versus Opus 4.7 5.5 scores about a 62 and 4.7 scores about a 33. So there's 30 point difference. But the really interesting thing is that 5.5 scores the best. It gets that 62 when it uses an Opus level plan. When it uses a plan literally written by Opus 4.7. So that's really interesting right? And I sort of tried to tease apart what what was going on. So the thing that 5.5 can do that 4.7 cannot do is if you give it a prompt that says hey I want you to like go and rewrite a major part of this code base. I don't I don't know what it is. But this is just a vibe this is a vibe coded slop code base. Like figure out how to how you would rewrite it from first principles and then do it. Both 4.7 and 5.5 can figure out okay here are the core principles. Here are the core invariants that I would rewrite and write a plan for it. But what happens is when you then say okay go execute it Opus 4.7 says to itself says to you like hey actually like this is way too big of a project. I'm just going to pick off like a little piece. And even if you kind of push it a little bit to like no do the whole thing it will still only do a little bit of a patch over the problem. It doesn't want to go and like actually rewrite the whole thing and actually delete a bunch of code. So it sort of gets distracted and a little bit almost intimidated by a really big code base and a big rewrite. Whereas GPT 5.5 is really interesting for its ability to take that plan and actually execute it over many turns over many many many hours over many many tokens. And for having the sort of like ability to to like have the almost like the courage or the assertiveness to be like okay I'm going to go and like delete a bunch of code and I'm really going to think about this from first principles without getting as distracted by the existing code. And that's just something that that's new. Like I just have not seen that in in an Opus model. I've not seen that in GPT5.4. It's even less present in GPT5.5 on high reasoning. This is only really on extra high. Um and and some of the things when you think about when you when I looked at, okay, it's it happens on extra high and it hap- happens on extra high with an Opus 47 plan. What's different about the Opus 47 plan? The Opus 47 plan is is very it's very terse and very spec-like and very contract-heavy. It says things like a good rewrite will take this gigantic file and get it down to 500 lines. Like it has almost like contracts that it's giving to 5.5. And even though Opus won't itself execute those contracts, 5.5 is totally happy to go do that. And I think that's a really that's a show something really interesting about the character of this model. And I think it's part of what you're seeing, Naveen, in in your vibe coding results because you're giving it you're not doing a detailed engineering spec or whatever, but you are you know, if you go look at the original prompt, you are giving it a lot of detail that a a like baby vibe coder maybe wouldn't. And it what it's able to do is take that detail and keep it in its head and actually like turn it into like a full thing um from start to finish. Does that make sense? I'm Naveen, by the way. I see that same result. \>\> \[laughter\] \>\> Yeah, do you want to talk a little bit about \>\> I would I would love that. Yeah. Yeah, so I I have a um uh one of my benchmarks is is like purposely very underspecified. Like I don't use plan mode and it's just uh create a version uh like a clone of Typeform uh but call it Talkform uh and the goal is like the back end is exactly the same as Typeform, the structured data, but the front end is you just talking to the person, you're interviewing them to get the answers for that structured data cuz it's like I don't think this exists uh but I think it'd be cool if it did exist and I've vibe coded that a few times in workshops. And um and and the prompt is like four lines uh on purpose uh because I what I really like to see is like from scratch with no repo, with a very underspecified prompt, like as a beginner, like is this model accessible? And uh so I we we talked about this because I was like, "Oh, it didn't do very good on vibe coding." and then you guys are like, "What are you talking about? This is the best vibe coding model ever." and I think that was the difference. Um uh I think Opus is like very ambitious to start uh and then it has like I don't know if you've noticed it says like, "Okay, are you guys ready to wrap this up?" You know, like if I'm having a long conversation with it, it almost like feels like panics that it's using too many tokens. I don't know, I hate to anthropomorphize, but uh that's how it feels. It's like, "Okay, do you want to wrap this up?" and I'm like, "No, no, not yet." and it's like, "Okay, but now do you want to wrap it up?" And and so I I wonder if that's the difference between uh what we're doing. Whereas Codex seems pretty chill, like it's like dogged and determined. It's just going to like keep pounding it out, you know, time after time after time. So I think like it really depends on what shape of job that you're doing uh when you're using these models. 100%. Naveen, any any reactions there? Um yeah, I I I I don't completely like, you know, I'm not seeing that, so maybe it's just that I do a lot of big brain dump prompts and then recently uh like I'm with Dan what you observed, which is uh I recently had to add MCP support remote MCP support to Monolog. I initially implemented really bad MCP server. I think uh I just did it like uh so what happened is last week we are like, you know, doing a launch. I just asked GPT5 5.5 to come up with a plan. This is where I used the plan and it added a plan across front end, back end, client side, both iOS, macOS, right? And then I just said I'm like too lazy, so I said, "Implement this plan." And then it did that. Like it continuously kept on working across all multiple code bases. And that's when I know, okay. Uh this model is really good at managing that multiple code like, you know, across context. I think that's why I'm like really curious to test this single thread uh vibe coding app as well because it kind of like uh I'm not sure what they're doing with compaction. I think the context between compaction or the context between sesh each thread that models are like, you know, using is really good. So it's not losing that context. I think the product shape that uh you know, I think that's what makes this model really good for me. Yeah, I actually have like a side project where I have a one of the Karpathy-style knowledge bases uh that's that I'm doing it away from keyboard, like it's just running in the background. And I I initially set it up as a Ralph Wiggum loop, right? Like uh every every time code stops, it uh you know, it checks uh to see uh you know, like what's the next task you have to do and then uh it takes that task down and then it commits and then it stops and then it starts up again. And uh and I switched to Codex uh on on that uh a while back. And um and and uh after it's just been running uh and and uh I I realized that I'm like it doesn't need that uh that like Ralph Wiggum loop anymore, so I just switched it basically to just let it compact itself. And and uh it's been going a lot faster. Like I need I needed less harness essentially, which was pretty nice. Um do you I'm I'm curious, Mike, do you have examples to show us of of more knowledge-y type tasks that you worked on it with? Yeah, yeah, I can pull some stuff up. Let me grab my That'd be great. And then while he's doing that, Kieran, what else what else do you have to say about this model? What else do you think um you've you've experienced that we haven't talked about yet? Yeah, so one one observation was uh that design looked a little bit less good than the previous model. So um it it just looked a little bit chaotic, but uh design in sense of structure and typography looked better. So if you if you if you like typography, it's better. Um so so this is the app where I tried to just one shot from one prompt an entire app with like a rubber ducky store that where you can customize ducks and all of that. Like it nailed it. Everything worked, which is like this the first time with 4.7 this ever worked. So 4.7 and 5.5 are beasts. This is a web app. This is React, like a Next app. Um it's very very very good at at this. So if you do web and you uh need stuff to work in full detail, um this is great. Uh I've seen here, so I run the LFG bench, which is using compound engineering to uh just run all the steps by itself and it took three times more tokens in the planning and the review state, similar to 4.7. So it takes the tokens. This was extra high. But as you can see here, like yeah, it looks cool, the typography side and the alignment and the padding, but the the right side like is weird. Like what is that even? Like I don't even understand what it is. So there is some some degradation. Like the previous models were better at this in design uh and yeah, like I'm sure they're going back and forth. Like they're they're like, "Okay, now we need to structure more, make it less chaotic, uh more structured." Um this one is the island test and I said just make a cozy island and uh it's pretty good. Like it did the the the ground was blue instead of brown um and the animations were not uh working completely correctly, but it's very close. Like 4.7 Opus did better there. Like the but yeah, it's it's still like it's getting better every time and it's cool that you can see where they push the model towards. Um and yeah, so design-wise, this this one is actually good. Only the colors off here. Like the nice thing like what I've seen before with GPT models is like it adds stuff it doesn't really need to add and it does a little bit too much in the UI. That that's not there anymore. So like that extra, like it's it's more like Opus where it just shows the stuff it needs to show, which is which is nice. It doesn't do extra things. Um and yeah, so design, but like one shotting um like I did a I did a library and I just uh here I can share that I had an idea one morning where I like, "Oh, it's like actually one of the more important things nowadays are like doing a call with someone and going through a product and sharing what you like about it and not like about it." And um so I created this uh library called it's I shared my screen. Um but basically in the morning I was having coffee and I just pulled up this and and I was like, "Hey, what what if we have like a React package? If you can share my screen, then you can see the Uh got it. Yeah, so I created this uh like riff wreck as like, "Let's go riff together and record it." And what I did was I just record screen sharing and then feed it to a model to see what was told. But I was like, "Well, what if we record like the DOM interactions as well in addition to um all the other \[clears throat\] stuff?" So a model can very precisely see when you say oh this is dumb or if this should move somewhere else you click on something it has that element and then that can be extracted very well into like an improvement. Well as you can see it's three commits yesterday and it's basically initial commit one one review commit and one more review commit. I didn't do anything. This is just LFG in one go like one prompt in monologue copy into like compound engineering LFG flow into codex 5.0 in in 5.5 and it did it and it's super detailed like it it had oh yeah there was one iteration where I said hey like maybe we should have a consent screen it actually proposed that it was like hey like maybe we should do a consent screen that's probably something you want that's missing here came up in the reviews I was like yeah sounds good. Um but it has all these things and um basically it has these things hey you can enable recording for URLs by doing this and you can even um create your own program where you can do your own so like there's these details that are super nice that it adds that I haven't thought of but it's not too much it's not sloppy it feels neat and organized. Um so it's very good at these things this is typescript I use Ruby in Cora it's just not good at Ruby so it's just very hard for me to use because it's not good at Ruby it just does typescript stuff but then in Ruby. And that's not how you write Ruby so for me that is almost the biggest blocker like if you do typescript or anything open AI trained on I guess probably amazing but I built Cora and use Ruby and Ruby is not liking the model as much. So you can see like this is like an amazing experience and and I I rolled this out in Cora I will try it out basically it was just record your screen record all your events and you have get a zip file stored somewhere you can share it with me and then feed it and just say LFG ideates come up with the the things we need to build here and yeah that's just a morning one shot and that's the world we live in and I'm not even surprised like some people say oh wow wow amazing guys for me that's the norm one shotting things for me the next thing is like how can we one shot 10 things and release them automatically so I'm more like for me the bottleneck is not one shotting things because that's kind of working for me the bottleneck is like how do you collaborate as a human with the model because that's the unblock and I think um like the ideation the polish at the beginning and the end like that's where you need most help as a human still. I love it. Um thank you for sharing that Kieran. I'm going to now bring in our head of growth Austin. Austin welcome. What's up? Um and Austin has been testing this model for three weeks alongside us and he is in an interesting place because he's using it for head of growth type stuff so for more knowledge work type task and he's coming from Claude code. So Austin give us the headline on um I think you started this experiment using Claude code and co-work mostly and you ended using mostly codex. Tell us about your experience. Yeah over the last few weeks the codex app in particular and I think it's really because of like the the power and speed of the codex app relative to the Claude desktop app has become my daily driver for everything I do. For for someone like me without like a technical or engineering background I feel way more comfortable doing any kind of engineering work with the latest open AI models that can be everything from like like running code to build like underlying dashboards across multiple tools shipping landing pages that look at like our Figma design or whatever doing analysis creating plans and when I you first like were nudging me to at least try codex a few months ago Dan and I I did try it and I was like either I'm too stupid to use this it thinks I'm too stupid to use this or it's talking to me like like I can't hang at its level and I was like I I kind of wish I could get as much out of this as as as you did or as I knew like Naveen was a big fan of it and I just kept going back to Claude code in the CLI because I was like okay it's executing at a at like a level I like and it's it's like nice to me and and engaging and as someone who is still relatively new to this stuff that was important to me. Yeah you were like codex makes me feel like I'm dumb. Makes me feel like I'm dumb makes me feel bad about it and this came up a lot in the previous open AI models in codex because I would make a plan and in planning mode it would say like I have four questions for you and for each question for the options all the options I was like I have no idea what the hell you're saying please explain it to me and the tone of the open AI model was kind of like why? Why would I have to explain this to you? Why don't you just do what's recommended? And especially in knowledge work when I'm trying to run like a campaign or planning like it's a it's a weird feeling to have and all of that is gone for me in in the new model that we've been playing with for the past three weeks. I both understand what it's saying I also really trust what it's saying. The the best example I have here is we are about to launch our plus one product in a few weeks and we're sprinting to do it I'm sprinting to get a big like campaign together and I pointed in the codex app I pointed the model at our notion our slack where like all of our meeting notes and conversations have and I basically just like I did a monologue brain dump of like the high level version of what I wanted the goals of the campaign to be and a previous campaign plan we had and I said go make a campaign plan looking at this and it's informed by like my taste and the team's taste and thoughts because it looks at our meeting notes where we talked about what we want to do and the plan that we're going to go to market with is like 90% of what it came up with. I was like so impressed and it was so easy because it grabbed all of this context it asked me questions that made sense to clarify and the most important thing to me is that inside of the codex app it's so easy to use and it's just so \[ \_\_ \] fast. I find the Claude desktop app impossible to use due to the like clunkiness and slowness after having experienced how fast for knowledge work codex is with with these models. Um and then the the other thing that I really really like that I think is driven a bit more by how the codex app works of what I feel rather than the models itself is the way that it recommends automations to me for knowledge work and then just builds them and starts starts using them if I if I ask for them is so powerful and impressive and they're automations that I rely on every day so um to me like the only way I would switch back to Claude and Opus whatever new model as my daily driver is if the if the Claude desktop app got like dramatically faster and and better so as kind of like a an easy way to work it it was competitive with the codex app in in that way or if like something about the the model for knowledge work like far surpassed where the open AI models are and to me they're they feel about even for most stuff besides design I agree with Kieran that like we use AI for a lot of like video clipping image generation stuff and over and over again I find myself searching for Claude for that kind of stuff. Yeah that makes sense. I do I think it's really important just to like talk about how remarkable that shift has been like even three months ago definitely six months ago but even three months ago it was like it was one of those things where you just couldn't use codex for any kind of knowledge work. Um it was too slow and it treated you like \[ \_\_ \] And now it's super fast and it's just good for the kinds of things that if you're head of growth or you're like me you're doing like analysis of like some of our metrics or you're writing investor updates or whatever it's just like it's the thing that I want to reach for. I think obviously we've been talking about Opus 4.7 has some unique powers so for example I think it's planning is much better we think it's design is much better but it's also even in the places where it's roughly equivalent which I think there are a decent number of places like that it it is actually a lot slower. Um and 5.5 is just a fast model and that that makes a huge difference. I think speed is in some ways a it's it is a type of intelligence. Um and it gives gives you a lot more power to have it be so fast. Yeah it passes this like I don't have a benchmark version of me yet I kind of want to make one of like my like do my knowledge work while I'm in meetings test where I'm like that or like I did this on Sunday where I was watching the NBA playoffs and I'm mostly watching the game and I'm just like I'm like nudging it along and my brain is only maybe like 10% working and when I'm in between meetings I'll check the outcome of whatever it is and this is a lot of like planning execution stuff and it moves so fast that I like can't keep watching the basketball game or I can't like I like during the meeting I'm like oh I like want to keep go looking at it." And then it has something good that I can actually go do the work and I don't feel like I have that experience with Opus. That's really interesting. Okay, so we've got I have a question. Is it okay? Then? Yeah, go for it. Ask ask your Yeah, like how is the software ThreeMotion? I know you've been playing with ThreeMotion a lot, so how is GPT 5.5 with ThreeMotion? I have so a a thing that I do a lot as someone who like needs to make social assets um for us, but has no real design ability or sense is I like to point these models at designs our designers have already made primarily through the Figma MCP or to screen record how our GMs products already work. Like we launched a new version of Sparkle with these models you can do a screen recording, give it to give it to one of the models and say like, "Hey, make me a video with this storyboard." And um over and over again with the exact same prompts, the Open AI models kind of just like take the recording and like zoom in a little bit. It's like very very odd. It's like the laziest possible approach and with the exact same prompt, I find that Opus and like all of the recent Opus models have been this way for a few months. They take this level of creativity and they they engage with you back and forth and questions for how to make it more creative that like I haven't found it close at all. Um I'm intrigued to see how much Open AI catches up here because Dan and I were messing around with the new image generation tool for YouTube thumbnails and it was really good. And it's sick. It's so good. crazy. Um I was kind of hesitant to to try it because I I'm like in my head like this is the one thing Open AI models can't do. And um right before I jumped on, I knocked one out for the the the the thumbnail for this live stream was generated on the new Open AI image generation tool in like 3 minutes and I was like I don't have to give it any notes. See. Yeah, like GPT image tool is really cool. \>\> Can you show us, Austin? Yeah, I think so. Um let me \>\> Cuz I haven't seen it yet, so I want I want to see it. But yeah, it's like Let me see let me see if I can find the the chat. Uh so you can see like the conversation and just like all it took. Um Cool. All right, I've got it here. Let me share my screen. Um I would say All right, cool. Can you all see this? Uh yep, we can. All right, cool. So what I gave it was um two thumbnails actually that that Dan had made. With GPT image? With GPT image tool. So like it has it has this baseline. This is for the video we recorded before the um Can you just click on one of them? I just want to show people how good it is. \>\> Oh yeah, for sure. Yeah. So this is this is one that Dan made I think very quickly, right? I made this really quickly. This is like essentially one shot with uh the new GPT image model giving it a thought giving it a picture of me and this is a little bit more massaging with GPT 5.5 um and the new image model. And it's like that's that's kind that's just crazy that it does that. Um and it's very it's very good at like um oh, I want you to change this one little thing, but keep the rest the same. And it's also good at it looks like me, which a lot of these models have not been that's been harder for them. Yeah, it looks like you. This is like I would say like a 95% approximation of our logo, which sometimes these things are very bad at. So I gave it that and then I gave it a few more um just like screenshots of the video and basically said um hey, this is the text for the YouTube video we're about to do. Give me three riffs on this. And uh it made this one uh which is I believe what we're using. It also made this one, which is uh like a little busy, but really good. Like in the busy state, it's like quite quite good. Um typically these kinds of icons these things make would look absolutely horrible and they look pretty clean. And it has the the right number of of fingers and that's not from a picture. Like the fingers were not in there. That's crazy. Um and this was the this was the third one just off of that prompt I gave it. And the interesting thing for me with this is that my my process in um in Claude Code with Opus has previously been that I made a skill um for like a YouTube kit skill and I had to preload it with all this context where it has the YouTube API, it has the Figma MCP connection, it has all of this stuff so it can go look at the internal taste that we've created and what's worked for us previously and the analytics behind our YouTube thumbnails. And all this was was like, "Hey, we're making this video." And then it creates this thumbnail. And um you know, when we do these live streams and we do these vibe checks and reactions, we have to move so fast. We're doing so much stuff that um the only way we're able to react quickly with real insights and have stuff that performs well is tools like this that get us from zero to one in seconds. I love it. I love it. Okay, thank you Austin. Uh that's amazing. So we've got some really cool stuff coming up. We've got we're going to have Mike who's our head of consulting on our tech practice uh talk to us more about uh how he uses GPT 5.5 for knowledge work and his vibe check on that. Um the good, the bad, the ugly, all that kind of stuff. We're going to have some special guests joining, so there's more special guests, non-Every guests. Uh I will not say exactly who it is, but you'll want to stick around for that. Before we get there, what you should know is you got to stay hydrated. It's important to here. Mike, you got to stay hydrated. \>\> \[laughter\] \>\> It's important to keep yourself and your agent liquid cooled. You should also know that Every is the only subscription you need to stay at the edge of AI. And I want to show you a little bit about Every in case you don't know who we are and what we're about. We are Every. Um we do these vibe checks every time a new model drops. We've got one this is our vibe check for today. I want to show it to you. Look at this, it's beautiful. We've got a full team across coding, writing, design. We we build tons of apps. We're we're you know, we're we're building it when using Open Claw. And what happens is we get access to these models. So we got it we got access to 5.5 about 3 weeks ago and we get to put it through its paces. And so you get the benefit of if you're an Every subscriber of getting our take on day one from uh from our hands-on testing. A couple if you're just getting here um a couple things that uh are important to know about this model. We found it's it's really spikes on on some of its senior engineering ability especially in this in the sort of back-end uh department and especially if you give it a good plan. So on our senior engineer benchmark, which measures how good a model is at taking a vibe coded slop code base written by yours truly and rewriting that code base in the way a senior engineer would comparing it to real senior senior engineer code. And on that benchmark GPT 5.5 scored a 62 out of 100. By comparison, GPT Opus 4.7 scored a 33. So there's about a 30-point swing there. Um and so this model has some special characteristics that we've not found in any other model that we've tested. Um a couple other things that are that are really interesting about this model. It's it has that spiky senior engineer ability and it's also super fast and and and pretty good to talk to, which is a it's rare to see a model improve on both dimensions at once. I think you can you can see that a little bit with um with Opus 4.7. It got a lot better at programming. It's it and it but it's also much more terse and a little bit more like it feels like it lost some of its empathy. Um and I think Open AI did a did a really fantastic job of making this much smarter, but also keeping it pretty pretty personable and pretty fast so it's usable for the kind of like day-to-day knowledge knowledge work tasks that that we're that we're used to. When we at Every when we do these um these vibe checks, we always do a reach test, which is are you using this every day? Are you reaching for it over other things? And if so, what are you reaching for it? Uh I'm a I'm a green on this model. This is my daily driver for pretty much everything. The only thing that I still use Opus 4 is every once in a while when I'm not talking to my claw, I'll I'll talk to uh Claude on mobile. I think it's I I really like Claude on mobile. But other than that, it's it's pretty much my go-to. For Kieran, Kieran is a yellow. I think there's some things Kieran you like, some things that you don't like, but it's it seems like Opus 4.7 is still really your your go-to model, your your daily driver with using this this model for certain other things. Naveen, uh you are a green. You're usually an Open AI stand, so uh so you know, know that, but I think that you you just you love this you love this model. Um Mike is a yellow and uh Katie Parrott who's our uh lead staff writer is a green. So it's not all greens, but there's there's a lot of people excited about this model and even if you're a yellow, I think I think pretty much everyone agrees there's a lot of power here. Um a couple things that we so that we saw. So again like it's really good at doing that senior engineer benchmark. Um it's really good at uh if you give it an idea for a a big thing it's going to build like pushing that idea through over many rounds and many many tokens, um it's a little bit weak it's it's really good at some of visual design stuff but it's a little bit weaker at other things. So for example, Kieran did this uh benchmark test where he asked it to to design this Kieran, do you want to give us just a little bit of a uh refresher on exactly what we're seeing here and what you think it says about the character of GPT 5.5 as a model? Yeah, and it's also good to understand what it is the benchmark because there I also test the harness here. So like it is testing the the the Codex CLI harness and it's testing the plugin framework and it's testing compound engineering. So there is a lot of things going on here. It's it's not just a prompt in uh result out. So it's also like how does GPT 5.5 work within the Codex harness? Like what kind of features does the Codex harness have? And that is also represented in here. But what I've seen change from previous things for design is like it's better at typography. It's less fiddly like weird details that were in in the the previous version but it's not as good at like visual. So you can see that the the left side looks good with the typography but the right side looks it doesn't doesn't make any sense. Like there's like there does not really make a lot of sense what it even is or what it's trying to do there and that was better in other models but also I do see it improve in structure. And maybe you can argue actually structure is more important and finishing things and not having any bugs and like staying on course is more important because these visuals are probably done by a designer or by an image model which they have which is very good. So like you could even argue that's I'm not saying this is bad or good. I'm just observing this changed and normally in these changes if you see them in different places you can see what they try to do. And this is a great model but also I test the model within the harness. So I when I say yellow it means yellow for also the harness in which I use it. And the harness is is very good. Um but there are other harness that's also very good. So also what what I'm what I'm what I like to do is I also recently like to try models against each other in the same harness in like cursor for example and and just see them head to head. I will do more of that. I remember last time I did a vibe check for cursor I say yellow. And now I'm using cursor every day. So like me saying yellow on 5.5 means maybe that I need to adjust how I work or learn the powers or it just means I need to change something to see it work and I'm not saying I won't try it. I'm actually running right now I'm running like three Codex things in the background doing stuff. So like I'm using this model for sure. And but it's a big thing to say daily driver and go super green. So like I'm very very enthusiastic about this model. It's super good and I'm very excited that I can do things in one shot because that was never really the case with like older versions where it just stop working or got trapped at some point. I think 5.3 was the first time where I was like okay OpenAI you got my attention. This is actually like very very very usable and 5.5 is just that but then way better. Like it's like if they keep going this is great. This is the best thing any engineer or creator can have. We're just like they're so good now that like yeah, it's hard to compete and they're so good and you can all use all the models. So really the vibe check here is for me like what is different and what are the differences between all the models. It doesn't mean it's good or bad. I love it. Um okay, so we've got about 3 minutes before our next guest Mike Taylor our head of technology consulting at Every. Do you want to tell us you've been using this for some of some knowledge work use cases? Do you want to give us some some thoughts on on what you found in doing that? I'm just going to open open up some of the things that you probably going to want to be looking at. \>\> Yeah, do you want to open the dashboards? That's the most surprising. \>\> yeah, let's do let's we'll do the dashboards. Okay. Yeah. So this is a Opus dashboard and you'll probably recognize that Well, maybe like introduce the benchmark first. Yeah. \>\> Yeah. So the benchmark here is this is actually something that comes up again and again in every workshop I do. Can we make a custom dashboard from it? And this is something that used to take you know a lot of time a lot of IT tickets to accomplish and now people are just vibing it. And this looks vibed, right? This is Opus and you can see like you probably recognize the color scheme if you've done a bunch of vibe coded dashboards. Like if you just scroll down kind of you've got like uh yeah, it's like very very bright like dark dark background and you got the purple at the bottom as well the vibe coded purple. \>\> Oh yeah, the purple \>\> \[laughter\] \>\> always. Yeah. Now if you switch over to the other dashboard so this this was 5.5. This is GPT 5.5. Right? So I mean it's like way better. \>\> It looks more professional. \>\> It looks more professional. Like if you scroll down it starts to get really nice. You know the color scheme was like more tasteful. The the spacing was better. Actually like wanted to read it. So I would say like this is the one this is the first benchmark around I was like oh my god this is like AGI. And and it's like it's a little bit more nuanced after after that but just wanted to point that one out because that's a like a comes up in every workshop and I would say like this is like a very very clear win and it surprised me because I I just saw that like with Opus did the best PowerPoint deck I've ever seen and the reason was that it was better at design and it had better vision than previous models. So I wasn't expecting this. Weirdly you know GPT 5.5 failed the PowerPoint test versus Opus but it beat Opus on the dashboard. So it's not a clear picture that it's like worse at design better at design. I think depends on what you're designing. I love it. Thank you for thank you for sharing that. We have now I'm very excited to welcome to the stage a very special guest Romain. Are you ready are you ready to be on the stage? No, it's going to be a little bit longer. Oh, cuz you are. I'll take you off if you're not. Romain is not ready. Okay. So just a just a little heads up. Romain Hewitt is the head of developer experience at OpenAI and he's joining us to talk to us about this model. That's coming up soon. And while we wait I'm going to share my screen again and I'm going to just say Every is the only subscription. Every is the only subscription that you need to stay at the edge of AI. We get these models way before they're released and on release day you can expect a vibe check like this from us where we use it for everything from coding to writing to design to marketing to strategy. And this page is not loading but this page does load. We've got a really really sweet long vibe check here. It's several thousand words. A lot of it's free. A lot of it some of it's for paid subscribers but if you're an Every subscriber you get a lot of other cool stuff. So as part of Every we have a bundle of apps that we build. Some of the apps are built by some of the amazing people on this call on this live stream. We've got Cora which is an AI agent for your email and soon an inbox. Kieran I would love for you to show us some some inbox type stuff if you feel ready to do that. We also have Monologue built by Naveen. It's a smart dictation app and it we recently launched Monologue notes. So Monologue started as it's like a speech to text app sort of like Whisper for Super Whisper but we just added notes. So you can transcribe a note on on your walk. You can you can record a note in a meeting. You can brain dump into it. It's really great. I love it and it's very available to all your AI agents. So it's really easy to get your notes where you want them. So we've got Cora. We've got Monologue. We've got Sparkle which organizes your desktop. We've got Spiral which is a agentic ghostwriter. We've got Proof which is a markdown editor for your agent and we've got Plus Ones which is our hosted open cloth. So this is all available for Every subscribers. You pay one price you get access to everything that we write, everything that we make and all of the we do a lot of we do a lot of live streams and and camps and stuff to to teach you how we how we use stuff at the edge. And we also have a pretty cool consulting practice where we work with big companies, work with executive leadership teams to get their hands on get their hands on stuff like cloud code or Codex and get them using agents and get agents into the org. So if you have not checked out Every you should go to every.to and you should subscribe. And I think we're still waiting I think we're still waiting on Romain. While we're doing that Mike anything you want to add on some of the benchmarks you did or or maybe anything I haven't talked about on the consulting side? Yeah, so maybe if you want I'm going up those two files called persona one for each model. So I'll tell you a little bit about this benchmark. So last year I worked on a startup failed unfortunately but which is why I'm working for a living. But but I mean Your loss is my gain. \>\> \[laughter\] \>\> I'm still going as a side project here. We have we have some of customers and stuff but but but it's now largely away from keyboard have Claude and Codex working on it while I'm here. But uh Yeah, it's it was a virtual focus group startup. So we used AI to role play as your customers. So I got very very deep last year into cloning people. Like I you know, we built a whole panel of 300 people that we cloned based on user interviews. And then this is my clone just for ethical concerns don't share anyone else's clone. But this is this is a clone of me just a snapshot taken just before I moved here to New York. And yeah, if you're seeing that on the screen here. It's just a great benchmark because obviously like We have spiral right where it's like it learns to write like you. But the way that I write isn't the way that I talk and especially once it's gone through editorial it's a little bit different. And I think it's maybe an easier benchmark to to to like write like an editorialized piece of content because that's more consistent is more standard. The way I talk is is really odd. It has like inflections. It has a lot more personal information. So what this is is I I provided with a transcript of this long half an hour interview I did. And that's the prompt. And then I have some extra stuff on top to say basically impersonate me. And actually the first time I tried it it refused to impersonate me. \>\> \[laughter\] \>\> I think some of the checks Alignment is is good and well but so so the first version failed in this benchmark. But that that was quickly fixed and and it worked beautifully. So I don't know if you have the this is the Opus one is it? Yeah, the Opus one if you if you look especially if you scroll down to like maybe the second or third question. It's it's it's trying too too hard. Like there's a lot of like ums and ahs I do me a lot actually like I think. \>\> \[laughter\] \>\> Exactly. Maybe I do talk like I mean you I like there's some some of that but not not all of that. Yeah. Yeah, and I think it is just like feels so eager to be me that it it says almost like a super version of me. Like it's really emphasizing my ticks. Whereas if you go to GPT 5.5 it's just much cleaner. Um It's just like a much smoother experience. And it picked out a bunch of things that I actually really do think or or like very very plausibly like could think which which is like as good as you can get really for this type of benchmark. Great. So I agree. I think we found the writing to just be a little bit smoother. Good at good at voice but not overdoing it maybe a little bit less AI-ish. And now I am very excited to welcome Romain Huet and Dominic. Dominic how do you say your How do you pronounce your last name? Kunda but you can just say Dom it's fine. Okay, I'll say I'll say Dom. Welcome both of you. Can you introduce yourselves and \>\> Yes, thank you so much for having us. This is Romain. I lead developer experience and I found Dominic joined me. He leads a lot of the DX efforts for Codex. And so we're both very excited to talk to you today. Amazing. We're psyched. We've been we've been using this model for the last 3 weeks. We love it. It does some special things that we've never seen a model do before. What like tell us about like what what you've seen internally and what you're excited about from this model. I mean it's pretty amazing how much now like the work has changed at OpenAI like relying on this model. I think we we've seen like across the board how much the model is smarter at any kind of professional task, right? Like we talk a lot about writing software and of course we we hear from engineers that the the model is writing better code. It's like more thorough when it comes to like analyzing a repository of of complex code bases but also takes on like more and more of the other tasks that we do every single day. And and I think what's very exciting to me is not just the model itself like GPT 5.5 but it's also the combo with Codex. Because when you have such a great model you also have to have the right software and environment to to make it shine. And and we're seeing like so much of that with Codex, right? Because Codex can help the model not just like figure out the files in the context but also access all of the tools that it needs to really carry the work checking some output and really like deliver things that you can just ship, right? Yeah, I mean like I I I feel like 5.5 has especially in a combination with like the features we shipped last week in the Codex app and some of the features we shipped today has drastically changed how I work like compared to even just a few weeks ago. Like I know that I can have Codex handle everything from like the documentation to like I think I built like 50 demo apps in the last week just sort of like trying to push Codex to its limits and like like there's surely no way that it can do it and it just sort of like hammers through it like last night. I think it was like 11:00 p.m. or something. I was like let me see if I can like one shot like a full 3D modeling software. And like it it fully did it. Like I I remember one of our colleagues was like well can you add like download like support for like the file type for like 3D files. I was like it already had it. Like it was already there. So like the level of like attention to detail that like the model puts in is is incredible. Dom, do you have any of those demos that you can show? I don't have the 3D modeling one but I do have some. Let me share my screen here. Sweet. Very prompt demo by the way. So we we just got out of the launch room flipping all the switches publishing the blog post and we just showed up. So Amazing. I we're doing it live. We're doing it live. \>\> That's what we like. How we roll. We want it we want it like this. Okay. Dom, you're live. We've got your screen. This was one of the one of the moments where I truly had sort of like a you know, mind-blowing moment. And it like the final result is actually in the blog post as well if you want to take a look at it later. But my colleague sent me this picture, right? Like this is just a screenshot off some software that he had been building. And it's mostly like for like illustrative purposes. So I thought like okay, let's take this and I gave it to um 5.5 and I simply told it like you know, I had a couple of other folders in the project that like I didn't want it to deal with. So I told it ignore those but then use WebGL to and like the real data from the Artemis 2 mission to actually implement this. And and it went ahead and like fully built this as a working app. And it \[snorts\] went to the um you know, it it cited its sources. It actually went to NASA downloaded like the flight data of the Artemis 2 mission and then rendered all of this. And some of the attention to detail that originally I was thrown off by was um it had um If I open this in our built-in browser here, it had this like slightly weird curve where I was like why is this why like this doesn't look fully like how the data like how the curves looked when I saw like animations from other people. And the result was actually that like it saw in my image that I had given it that these two planets were like you know scaled like this. So it tried to like replicate the scale but that's not real scale in space. So it had to like take the real data trajectory and then like fix it for those for those moments. I I asked it later on to like add a true scale moment. So like here you can see the actual sort of like slingshot and then how it goes over to the moon. And if we go Towards the end is also a little fly by which is really We have like the full fly by and like how it like sends the moon around it. And like all of the data like it tells me here like what were like the mission moments and stuff. Like this was like one of those moments where I'm like I I didn't give it a lot of instructions on what I wanted this to look like. I just wanted it to Look at the picture. Look like the picture and it was like cool I got you. And so I've been experimenting a lot with this. And so one of the things that we sort of silently launched because I definitely wanted you to try it but like I didn't like we just sort of like shipped this is inside our like built If I find it. Move web apps. Built web apps plugin if you have that already installed or if you install it now, we added a new skill here called front end app builder. And this one actually uses now the the design capabilities of ImageGen that we launched this week together with like 5.5's attention to detail to build apps for you. So I told it earlier like right before this like you know, GPT 5 is out. Here's the blog post. Build like a minimalist well-designed page. And \[snorts\] it went in and like you know, this is one shot. Like the other things is me just asking to start the app. But this is the app that it built. Wow. With like the data from the from the website and the design if you look at it is like pretty spot on with you know, minor details of like ImageGen generating the individual assets later on inside of different positions. But like it's pretty spot on for like a one shot. That's really interesting. So is this like is this a new workflow for doing front end design is like have the image model like render a website and then have 5.5 turn that into HTML and CSS. Yeah, I mean it's a great one. I mean we've heard like, you know, from the community that one of the things that we still have to improve and push the frontier on is really like the front-end and the taste kind of capabilities. But now like and we're still pushing on that very hard. But at the same time like with Imagine 2.0 the quality of the designs that you can create is pretty outstanding. But also GPT 5.5 is so thorough at taking an image input and implementing it that you can now iterate on your design first and when you're happy with it just get it done with GPT 5.5 to implement it like this one which is pretty magical. Yeah, and like I think I think one of the interesting things there is like it really brings together a lot of the capabilities that the model has gotten better at like one like the attention to detail but then also like it went through the entire like like loop here. So like it used one of the things we added today as well as browser use inside this in-app browser. So it actually controlled the in-app browser to like take screenshots, try things out. Um and then it would like do checks here on like, you know, does it work on mobile? Does it work on desktop? Take screenshots and like verify that work and really put in sort of that attention to detail before it felt comfortable with like, all right, this is good enough now. Um And that's really the magic of bringing like the Codex harness, the Codex surface and all of these tools, right? Like I think if the model does not get everything right from the first try you don't even have to interrupt it because it checks it won't work. Like now with computer use outside of the Codex app but even inside the Codex app you get the model to just like keep on iterating until it's satisfied with the answer. Hm, that's really interesting. So what I what I'm really curious about is I feel like you guys are on this crazy arc, this crazy run right now where if I think about Codex 3 months ago or 6 months certainly 6 months ago it was like very there was no desktop app and it was very like senior engineery both in terms of the the user base and what it was like to talk to. It was like super slow. It would get very very technical answers. Austin who's our head of growth was on was saying like it just made me feel stupid. Um and I feel I feel like over the last especially the last 3 months Kieran was saying since 5.3 which I agree with you've been on this run where like every release it just it's gotten smarter, it's gotten more personable and and really good for senior engineering work. Like it's it it tops our senior engineer benchmark for what it's able to achieve but it's also just really good for just general agentic knowledge work. So tell us about that shift and and the trajectory you're on. Yeah, I mean like it's it's almost something that has been like very surprising to us in a great way. Like it's the like even at Open AI internally the adoption of the Codex app, right? Like it started with us being engineers and building with it and of course as you know we kind of started with the cloud and then the CLI and then the IDE extension. So we kind of had like different manifestations of the coding agents. But we always believed into this like agentic delegation vision and we truly needed a surface with the Codex app to like achieve that that that vision of like being able to send complex tasks and have the model handle them. But what's happened in the past like few weeks even at Open AI and outside of Open AI is like not just the best engineers in the world like moving to Codex and taking on this like agentic delegation workflow. It's also more and more of the like adjacent teams, adjacent roles like actually using the Codex app for any parts of their work. I mean like Dom and I have like I've completely changed the way we work at Open AI. Like most of us here like not just for building software but also to connect to like your Slack and your Notion. We like you can show the plugin screen but we have now like like 100 plugins where you can connect like all of your tools and all of the contacts to bring into this Codex app. And as of today also with the launches we also introduced like the artifacts. So the ability to like, hey, if you're working with spreadsheets or with if you're working with like any kind of documents and files or if you produce like assets you can now visualize those in the Codex app. So it's really turning into a very advanced kind of like productivity app for much more than just a developer. Yeah. Go for it. I was just going to say like I think one of the interesting things is that like um as I mentioned like over the last couple of weeks like the way I work has drastically changed. Like with these launches like especially we had a big launch last week. We had a big have a big launch this week and we had smaller ones along the way, right? Like, you know, Imagine was certainly a big launch but then we had things like Chronicle launch in in research preview which inside Codex like can watch your screen so you can learn more about how you're using other tools. And all of these sort of different capabilities have really helped me like stay on top of all of these things where like the amount of times that I would kick off a task because of some feedback and I wouldn't even actually go in and you know describe in like long words what I need to get fixed out. Like often a prompt would look much more like, you know, um like go like go look at like the last feedback that like Roman sent me on Slack and like go fix that and send him a send him a message when the PR is up. And like that's the that's the full prompt and 5.5 is intelligent enough to like figure out with like my memories and my context window like which Roman in my like Chronicle history like which Roman am I talking about? It knows to use Slack to like pull that message. It then like knows possibly like what are the docs that are related to the to the launch and then actually like what uses an automation to like keep it keep track of the PR until the deploy preview is live and Slack it back to him. And like I I just send off one message and like it's wild for me to think that like it can just do that in like a message. \>\> Yeah, and I think what's really magical too with the launch of last week to put things in perspective is that now you have the Codex app but the Codex coding agent historically could only manipulate like your files in the repo for instance and then connect to some skills or plugins to kind of bring that context. But there's a lot of work you do on your computer that is not always just an API or a plugin. And now if you combine like Chronicle as a research preview we kind of like which kind of understands a little bit how you work on your computer but also computer use. I don't know if like you all watching this I've I've tried it but I really recommend trying it because you can't really describe until you experience it like how good it is because to my knowledge there is no other computer use implementation out there where you can simply ask a task to Codex and then it would kick off its own like app in the background and then its own cursor and you can still use your computer. And so you can still like send like multiple of these tasks in the background for those that don't have let's say an API or a plugin the model can actually figure everything's out including your your native app on a Mac. Let's say like the reminders app or the notes app. Like anything that you use on your computer Codex can also drive. So now if you bring all of these things together you have this like combo of Codex and GPT 5.5 to really drive any task that you do in your days which is pretty pretty outstanding. Really cool. Kieran uh ask question. Yeah, I I would love to like it's it's really cool these examples and this inspires me to like actually like really kick the tires and clearly you know the model better than me since you used it even more than we did. Um Um one question is with these examples which are very cool like can you also if you are allowed to share more concrete examples of like what you do in like actual day-to-day work and how you use 5.5? Because yes, this is a cool example but like do you ship Codex or do you like use harness engineering or like maybe maybe harness engineering changed with 5.5. Like I would love to hear some of the like the inside stuff that you can share around these. Yeah, I mean I think one of the I'm like trying trying to think back of like the last couple of weeks of like really leveraging 5.5 what are like some of the standout um examples but um I think one of the one of the parts is right now I've spent a lot of time sort of improving um like the Codex documentation developers.openai.com was like one of one of those examples and um we had this on what? Like Sunday night before we launched Chronicle the next morning where I thought it would be helpful for people to understand sort of the impact that Chronicle has by looking at um like having side by sides. And so we have on the regular openai.com website uh the um like a mini Codex fake Codex component. I had never used that. I didn't know where it was, how it worked. So I just like told Codex like go look at like what how that's implemented on the website, pull it into the developers website which is an entirely different code base, implement it and then like read through the docs that I had drafted in a Google Doc to figure out how it actually has to like document these examples and like create them. And all of those things were just sort of like running in the background. I think like one of the interesting things is that like I can contribute way more features now even sort of like smaller delightful features to the app or like try out an idea in the app. Not all of them necessarily ship, but sort of like inspire people inside the team um without necessarily doing a lot of context switching, if that makes sense. And I was chatting like on top of that I was chatting with some early developers like you done who have had access to the models to for like a couple weeks and like the things that that came coming back in those chats were like I am not even specifying like much of the task anymore. I just trust the model much more to figure things out. Like navigate large complex code bases and like extract what's needed, solve complex bugs that were going to still hanging around, but now like 5.5 could solve them. And to your question Kieran on on harness engineering, I was chatting with Will at an engineer at Ramp. And they're using their own harness, so I was kind of curious how that worked for them. And he told me actually like complete plug and play. Like from GPT 5.4 to 5.5, but better yet 5.5 started to discover new tools that other models could not discover. It's like all of the sudden it it realized how to access the database, how to fetch data. It was complete plug and play, but also like new magical powers they did not have with any other model before. I love it. Thank you guys for sharing. I know you only have like a minute left, but a big question that we have not gotten to yet is GPT 5.5 is out right now in Codex and in ChatGPT, but it is not available in the API. Can you talk about why and when we can expect to be able to use it in the apps we're building? Yeah. We are taking like a very you know safe approach for for this launch and we are very very eager to give like 5.5 to everyone. And so we expect that to come extremely soon. Hopefully in days maximum, you know, we're we're just like making sure that that like the roll out goes well and that we have all of the safeguard in place for for everyone to benefit from this model, but again back to our OpenAI mission, we want to make sure that everyone can put their hands on this model. Awesome. Thank you guys so much for joining. It was it was a pleasure chat with you and and thank you for the model. It's great. Thank you for having us. \>\> Thank you for having us enjoy GPT 5.5. Of course. Thanks. See you guys. All right, so that was the OpenAI team. That was pretty cool. Kieran, Naveen, Mike, any thoughts? Anything you want to share? And Laura, welcome. Laura is a staff writer at Every. Laura, hopefully you're ready to be to be live because you are live. Welcome. Thank you. I'm not ready at all, but thank you. \>\> ready. I will take you off then. I'll take you off. Um so Mike, Kieran, Naveen, any any reactions to to the the OpenAI team? This is why I come in the room dancing. Are you confident? \>\> \[laughter\] \>\> Yeah, I I I I really want to see more computer use stuff. I think that you know the Chrome extension for Claude is like one of the main reasons I use the model still as well. So I haven't had a chance to really kick the tires on on that, but it's also the thing that all of my enterprise clients are most scared of. \>\> Yeah. So it'll be really interesting to see how that rolls out. \>\> Which is an interesting sign like if people are scared of it, it might actually be the most useful thing. Like I feel like one of the things that held OpenAI back for a while is that they were super scared of unhobbling the model and just like letting it do stuff on your computer. And now it's just like oh yeah, it just does stuff. Um So that that's really interesting. I love I love this. Okay, you're going to do it image generation with this sick image model and then we're going to turn it into a front end thing cuz it totally short circuits the issue that that we have with okay, the taste of this model isn't as good and and so they don't even necessarily have to improve along the dimension of can it come up with a good designer on its own. They just like let the image model team do that and I think that's amazing. Yeah. Yeah, on that one it's it's kind of interesting because I have a benchmark where I have an image and I say generate the website for this image, but it was so bad. It just put that image as a background. And that's that's what 5.5 did. So I'm just like like I'm like oh Me too, but that's what most mo Are we sure it has thinking on? Cuz I've had that happen, but I didn't have Yes. Yes, extra hot. Yeah, it's on, but like it means that it's not always the model and it's also the harness and how you prompt. So for me this is again like hey, can I rewrite that in a way or like improve the harness to see to get everything out of it? And and that's why it's so great to talk with the OpenAI because they have the best knowledge. Like we we played with it, so we can share our opinions and and findings, but also they if they can show you hey, this is possible, then then we can now fine-tune how we do it and learn from that. Which is like I would love to see this evolve into a daily driver by just learning how to use it. That's my take. I really want to try my PowerPoint work again now. The image Yeah, yeah, yeah, yeah, yeah. I I but I think the image following the image is something that I observed. It has improved a lot because in the past it didn't do it. I have couple of examples that I did recently. Let me share my screen. I was like quite shocked how well it's able to follow the images. So I'm had a opposite experience. So Daniel who is our board designer for Monita Log and Every, he gave me this design and this is the design. Oh, sorry. I think let me show you the old design that we have. This is the old design, right? This completely white coded. It didn't follow. And then this is the final design. Can you see any difference between the Daniel's design and this one? Wait, show us show it to us side by side if you can. So that's a screenshot of the actual developed actual app. Yeah. \>\> Yeah, you can see it's a local host one, right? Yeah, yeah, yeah. But uh show us the the original design that it was working from. \>\> Yeah. This is the original design. Oh, how can I Let's see if I can actually open this. Yeah, sorry. Maybe let me just put it there. \>\> \[snorts\] \>\> Yeah, I'm now I feeling good. So it just did Is this a one-shot thing or or what? Or so there is a good harness here. So I actually used it asked it to take screenshot, so it started the web server. It used Playwright browser. And then actually it it said browser binary is missing because I never actually used Codex for building web apps because I always thought it's bad at it. I reach out to Claude code for this kind of task, but I tested it. It actually downloaded the Playwright. I think somehow images are not loading here, but it just took ton of screenshots. You can see the whole thing is changed now. And then I started giving some feedback. It pushed it and then I gave some feedback and then finally I think it exists. So I gave a lot of like here it's not there. I really love the comments thing. Codex has this in-app comments on the UI directly. So I think I did couple of iterations to and fro, but finally it's like all done in off an hour or something. So that's very cool. Yeah. All right. It's like following the images really well. That's really cool. So if you just got here, we just had the OpenAI team on. They shared a lot about a lot about GPT 5.5. Um and uh and we have found that what they taught us is you can use the new image model to make front end designs and have GPT 5.5 implement them. I think that's really awesome. Now we have representing writers from all around the world. We've got Katie Parrott, staff writer at Every. Katie wrote the vibe check. So if if you saw the vibe check that we published today, which is around here on one of my tabs. Anyway, this is the vibe check. Look at that. Look at that. See, there's Katie Parrott is that's her name. That's because she wrote it. So Katie, thank you so much for writing this, for doing a lot of the testing. I would love for you to walk us through what you found when you used this product for writing. And I'm going to move over move our screen over to some of the stuff that some of the stuff you found. Yeah, totally. So just to set the scene a little bit, I have been I haven't touched OpenAI models for writing in almost a year. Around the time that I think it would have been Sonnet 3.5 dropped, I just found the writing quality on Claude models to be so much better that I just didn't feel the need to use ChatGPT. And in fact, I actually did what I thought was breaking up with chat with ChatGPT. I canceled my personal subscription about \>\> Wow. about 3 months ago. And then and honestly when I first started doing the tasks for this model I was like okay, nothing new here. I don't need to change anything. It really wasn't until I got my hands a little bit more dirty and like went back and forth with it and saw how it took direction that I was like wow, this is really good. It's So some of the things that we found and Mike can speak to his experience, too, is just that it's it's it's smoother. So it like the logical progression from idea to idea flows like flows more naturally. It it's less clever. Um I think that something that I found with Opus and if you scroll down to the third bright pink um you can't read it on the screen, but um yeah, if you go in there and zoom really close, you can see um that that this is the GPT 5.5 1. Um and um you know, at the end of Q1 I sat down to review my OKRs and discovered that I failed roughly half of them. Like I first read that I was like, well, you know, there's not really much there. So that's the kind of thing that I might want to like make more clever and more neurotic. But um you see and like and I've I've just told that like capture more of my neurotic energy, please and and it does that. Um I can't show you that because it might reveal some details about secret information about the model. Um But um yeah, and then like and and for me like a really big thing that is that's important to highlight is the speed because when you're writing with AI, you know, you're you're not going to get it's not going to one shot anything. You're going to go back and forth. You're going to be like this like I like the direction of this, but it's not quite right. Like this hook isn't the right hook. Um and so you want that speed of that and that ability to take feedback and go in the direction that you want with it. And I've just found that um GPT 5.5 is faster and better at picking up on what I mean when I give it direction even when that direction is like extremely um \>\> \[laughter\] \>\> extremely poorly phrased and very like minimal in terms of like, you know, direction, which is something that we found with Opus 4.7 is that it needs a lot of specificity in how you prompt it in order to, you know, follow your directions. So I find that GPT 5.5 to be a more intuitive model as well. \>\> So, are you breaking up with Claude? I am probably \[clears throat\] I am thinking about breaking up with Claude. Um \>\> \[laughter\] \>\> The beautiful thing The beautiful thing is because I work at Every I I have subscriptions to both. Thank you, Every. Um but for my personal use I will probably bring my subscription level down just because I don't see myself using it as much. Fascinating. Mike, what do you think? Yeah, for writing in particular I thought this is uh very dependable model. Uh like there's a lot of things that I write where I just need to get the job done and and I I just want to rely on it to not say something stupid. Uh and uh and it it does a great job of that. Uh I still prefer Opus for writing just because I like the neuroticism uh element like it it is it's very witty, right? Opus is very witty. And and um but the problem with Opus 4.7 I found is that it would say something clever and then just stop. And it's like very like I'm like I asked for a like a section and you've given me a paragraph. Uh and it's like almost like again that's anthropomorphizing the models, but but like it feels like it's like great, I've said something clever, now I can like move on to the next thing. Uh whereas uh you know, this is uh workhorse is is the good is a good term for it. I think you guys picked a really good title. Uh because uh it was probably Katie deserves the credit there, but uh but uh it is a workhorse. It it just like gets the job done without fuss. Like it just kind of churns through the work. And um you know, for most writing tasks I don't I don't need like neuroticisms. Uh I need just like, you know, job done. Katie, any uh any reactions to what Mike just said or any anything that you found in your testing that we have not talked about yet? Um I think um you know, I like I think like um \[clears throat\] Let me see. I don't know. Um \>\> \[laughter\] \>\> That's okay if if not. I can Yeah, no, I think that I mean that's more or less caught. It's so hard to talk about writing it's like a short of just reading out the prose that it produced. Um you know, it's you know, this doesn't make for good streaming. But um what I will say is that I did write the um the vibe check itself with um with Codex um with with GPT 5.5 and that was an awesome experience not just because of the model although although this may actually be about the model's capabilities, the Slack integration being able to pull information straight out of Slack and when Dan did what Dan does and zoomed in at like 5:00 p.m. the night before the review had to go out to flag an issue and talk it out, I was able to pull the additional messages, talk to to Codex about what needed to change in the in the draft, and get like new sections for Dan to be happy with, hopefully. It was great. Yeah, they came out decently. Um they they most of it made it into the into the doc um or into the published piece. And so that the workflow like having that really powerful model sitting at the middle of all these integrations um is something else, you know, speaking more to the general knowledge worker angle, which, you know, writing sort of nests inside of, you know, I just feel so much more resourceful sitting inside of this um this ecosystem, which now also includes G Chat GPT agents, which I am also extremely excited about. Um \>\> Have you tried? I've not tried yet. I so I built I've been working on a project to get AI to manage my OKRs for me because I failed at half of them last time and I was like, can open I like and I have an extreme weakness at project management. Like can't do timelines, can't do dependencies, can't do milestones, can't do any of that. So I just gave my OKRs to the model and was like, can you help me do this? And so like I um I basically yeah, I talked to the I had the agent I had an agent build me a project manager. Um and it yeah, it connects to my Notion task list, it connects to my calendar. So I can just say, "Hey, what should I do today?" and it will tell me what needs to ship, which is great for my brain and hopefully All right. Well, vibe check of the agents coming soon, but uh we are we are approaching time here, folks. We've been streaming for about an hour and a half. If you just got here, you should know Every is the only subscription you need to stay at the edge of AI. We have published a detailed several thousand word vibe check of our three weeks of testing written by Katie Parrott. Um the overall take that we have is this model is a beast. It's really good. It it it it is both a step change in senior engineering capabilities and it is quite good at just being a fast, personable like working collaborator, which is a pretty rare thing to like increase your efficacy along both of those dimensions at once in a model. Um it scored a 62 on our senior engineer which measures 62 out of 100, which measures how well a model how good a model is at cleaning up a vibe coded slop code base in a way that a human senior engineer would. It scored a 62, Opus 4.7 scored a 30 32 33-ish. Um so it's like a 30-point swing, which is pretty it's pretty crazy. One caveat. Uh the on the senior engineer bench the uh 5.5 scored the best when using Opus 4.7 plan. So um so Katie may be breaking up with Opus or or with Claude, but if you really want the best get the best out of your coding capabilities, I actually still think you kind of want both of these models in the mix. Um uh but anyway, this is really good model. If you're into it, you should definitely check out our vibe check at every every.to/vibecheck/gpt55 or you can just go to every.to and it's a link in the homepage. Um Every, like I said, Every is the only subscription you need to stay at the edge of AI. You should subscribe at every.to/subscribe. We do ideas, apps, and training. On the idea side, we have articles like this every single day. We have a daily newsletter where we keep track of everything that's going on in in AI. You just read one thing and you are absolutely caught up. You get all of the stuff that we're learning as we as we build companies with this, as we design, as we write, as we code. Um we also have a suite of apps that you get as part of the Every subscription. It's absolutely built for ourselves to help us do better work in AI um that we include for you in the subscription. So we have Spiral, which is your AI writing assistant with taste. We've got Cora, which is an AI email assistant. It's an it's an agent for your email and when it's coming out with an inbox soon, so you can I will be doing all my email through Cora very soon, which is great. Sparkle, which is a file organizer. Uh you can see my desktop is completely empty. That is technically my second desktop, but my actual desktop is also empty and that's because uh Sparkle organizes it for me. Uh we've got Monologue, which is a speech-to-text app. It's like super Whisper Whisper flow built by Naveen. Uh it's in it's really really good. It's super fast. We also just released notes for it, so you can uh you can record any note you want, a voice note on a walk, an idea storm, a brainstorm, your meetings. Uh they're all transcribed and all available for your agents. We've also got Proof. This is the app that I vibe coded a couple uh a month or two ago that we had to rebuild because it was the vibe code of slop, which became the senior engineer vibe code benchmark. Proof is sick. It is a markdown editor for your agent. It's all web-based, so it's really easy to share uh markdown documents between different agents or between you and your colleagues, so like coding plans and stuff like that. So you should definitely check out Proof, proofeditor.ai. Also PlusOnes. It's our hosted open claw. One click, it's in your Slack. All of this is available as part of your Every subscription. every.to/subscribe. All these apps plus all the articles we write. Plus we do a bunch of training. So we do live streams, we do camps, we do courses. Um we do all that stuff. Ideas, apps, and training all part of one subscription. We also have an enterprise offering. So, the the the base every subscription is really intended for individuals and teams, but uh if you're uh part of a part of a larger company, a big startup trying to trying to go AI native or agent native, uh we've got you covered. Mike, do you want to talk a little bit about the consulting offering? Yeah, sure, yeah. So, uh so I'm I'm a recent addition to the consulting team. Uh I just joined in February, but actually I've been writing for every for the past couple of years. And uh before I joined every, I like to joke that I kind of invented reinvented every from first principles. Like I was playing around with AI, I was writing about it, and then I was consulting. So, uh now I'm just doing it as part of a bigger team and more professionally. Uh as hopefully Natalia will tell you. \>\> \[laughter\] \>\> Uh it's been a change. I'm wearing a shirt for now, you know, I've I've had a haircut. No, it's uh it's not that corporate here. Uh no, but um yeah, in the in the consulting team, what we do primarily is uh AI transformation adoption. So, uh quite often um a company will come to us and say, "We just bought Cloud Code for this whole team. Uh how do we get them to use it? And how do we get them to be effective with it uh in a way that is like safe and matches like what we what we can do internally uh any kind of restrictions limitations." Uh so, it's uh you know, hands-on hands-on tactical stuff. Uh we're usually doing at least 50% of those uh workshops are uh like like actual actions. Like you're you're actually building stuff in the workshop. Uh so, that's uh that's a big focus for us. I think you can only really learn this stuff by playing around with it. Uh and we obviously have a little bit of theory uh to to go alongside that, too. Uh one of the things we've been working on more recently is uh trying to take the bundle concept, the membership concept uh from the the the rest of the business, and then thinking about what would uh subscription look like for consulting. Uh because time and materials doesn't really like make any sense uh anymore when uh you know, my agent can be working overnight on something for you. Uh so, uh that's something we if you're interested in that, like we're trying to explore that at the minute. We're talking to four or five companies uh about what that looks like. Uh so, uh yeah, happy to have that conversation. So, if you are uh if you want us to come train your executive leadership team, if you want us to come set up agents inside your company, you know where to find us. Otherwise, you should subscribe to every. And uh and if you don't, that's okay. We'll be back the next time there's a model release, so you'll definitely see us. Anytime there's a new model, check out every. We'll tell you what we think of it. Um it's it's always a pleasure to do these. I I \[ \_\_ \] love model release days, honestly. Um super fun. I never eat lunch because I'm too excited. Um and so it's like 3:45 here, so I'm going to probably eat some lunch. But, what you should know is you should always make sure to stay hydrated. Cheers. Keep your agent liquid cooled. Not hydrating. \>\> I saw that Coke Zero, Katie. That doesn't count. \>\> I was looking for my water bottle, and I couldn't find it. \>\> \[laughter\] \>\> Going to say you confused me. I got it. I'll be hydrated next time. All right. Thank you all for joining. Check out every, and we'll see you next time.