Dan Shipper, CEO of Every, shares his month-long test of OpenAI's GPT-5.6 Sol. The review covers the model's performance across coding, writing, design, and knowledge work; why he compares it to a Porsche; when to use it versus Fable 5 and Opus 4; a Babel bench demo; writing-quality and design-taste tests; and how models like this shift users from doing tasks themselves to managing systems that do the work, including email, meeting agents, and personal-life automations. Also noted: a new desktop app merge. # Video Identity Observations: "I Tested GPT‑5.6 Sol for a Month" ## Channel Self-Description (Explicit Claims) The creator states their identity and the channel's purpose directly in the video: - **Named identity**: *"I'm Dan Schipper. I am the co-founder and CEO of Every."* - **Channel value proposition**: *"This is the only subscription you need to stay on the cutting edge of AI."* - **Positioning metaphor**: *"You can think of us as a kind of advanced laboratory for the future of intellectual work."* - **Scope of testing claimed**: *"We spend all our time testing new models before they're released, testing them on everything from programming to brainpower, design, copywriting, marketing, and growth. We use this for everything within the company."* - **Content cadence and distribution**: *"On the day of release, we share these 'vibe checks' with you. They go on YouTube. And also appear on our website, every.TO."* ## Observable Presentation Patterns (Demonstrated, Not Claimed) - **Day-of-release review format**: The video is framed as a timely "vibe check" — *"Today is the release day of the new model"* — with structured evaluation categories announced: *"programming, writing, design, and intellectual work."* - **Comparative benchmarking**: The creator presents proprietary testing against competitor models (Fable, Opus 4.8), citing a *"Senior Engineer benchmark"* with specific scores (GPT‑5.6: 56/100; Fable: 91/100) and a self-designed test called *"Babel Bench."* - **Practitioner framing**: Claims of insider access — *"who am I and how did you get access to the model that came out a minute ago?"* — plus references to internal company usage (a colleague *"Austin"* using it for marketing emails; a personal tool called *"Tend"* for email management). - **Opinionated ranking system**: Uses a letter-grade framing — *"I wouldn't give this model an S level for programming. This is an A-level model."* - **Direct calls to action**: Repeated prompts to subscribe and visit the website — *"head over to every.TO, read the full Vibe review, and get the rest of the stuff included with your every subscription"* — and to engage: *"Leave feedback in the comments."* ## Intended Audience Evidence suggests the video addresses: - **Professionals using AI for work**, particularly knowledge workers: the creator emphasizes that the model enables a new approach to *"intellectual work"* — *"writing documents, searching for information, reports, chatting on Slack—everything you probably do at work."* - **Developers and technical users**: Detailed benchmark discussion, subagent workflows (*"I want you to use GPT‑5.6 as a subagent"*), and references to programmer practices. - **AI-tool buyers**: Repeated mentions of cost and credits (*"your Fable credits expire instantly"*, the model is *"relatively inexpensive"*) imply an audience making purchasing/triage decisions among AI tools. ## Claimed Scope vs. Demonstrated Scope - **Claimed**: Broad coverage — *"programming to brainpower, design, copywriting, marketing, and growth"* — for a subscription product. - **Demonstrated in this video**: Four categories only (programming, writing, design, intellectual work), with the deepest focus on knowledge-work automation and AI-assisted workflow design (*"instead of doing everything yourself, you control the system that does it"*). ## Notes on Source Quality - The transcript appears machine-generated with errors (e.g., *"CHCHT"*, *"CHBT Codeex"*, *"every.0/"*, *"Weibo"* for "vibe check"), so exact product/channel names (ChatGPT Codex, every.to) should be treated as inferred from garbled text rather than verbatim quotes. # Delivery Observation Report — "I Tested GPT-5.6 Sol for a Month" ## Overview A single-host review video (Dan Schipper, "Every") structured as a **"vibe check"** — a recurring evaluation format with fixed categories, scoring, and comparisons against rival models. Delivery is conversational, fast-paced, and metaphor-heavy. ## Recurring Structures & Teaching Habits ### The "vibe check" format as a frame The phrase recurs at the opening, middle, and close, functioning as a brand hook: > "So put on your glasses, grab your sunscreen, and get ready to check out the vibe." > "So, it was a check on the Weibo [vibe]." He announces the categories explicitly before beginning: > "There are several categories for this vibe check. This is programming, writing, design, and intellectual work." ### Rank/scoring language - Assigns letter grades: **"I wouldn't give this model an S level for programming. This is an A-level model."** - Cites numeric benchmark scores: **"GPT 5.6 scored 56 out of 100 in this test. Fable received a score of 91 out of 100."** - Names his own benchmark: **"what I call Babel Bench"** — an invented, repeatable test the audience can watch. ### Systematic comparison against a rival model Nearly every category follows a **"5.6 did X, Fable did better/simpler X"** structure: > "Same request. Look what Fable did. From the first time. Same image model." And a hybrid-workflow recommendation: > "I want you to use GVD 5.6 as a subagent... they seem to be made for each other." ### Central metaphor sustained over time The **Porsche analogy** is introduced, quoted from the model itself, and then elaborated: > "I perceive it as a Porsche... It looks elegant, not edgy. He is fast. He has excellent control." Contrast framing for the competitor: **"Fable feels like a really big model. It seems massive."** ## Tone & Vocabulary - **Enthusiastic, informal, superlative-heavy**: "gold standard," "the best model and the best environment," "world-changing genius of Fable." - **Playful humor / running joke**: The "Howard Latnick" government-shutdown bit appears twice ("Thank you, Howard Latnick... Thanks again, Howard") and again in grading ("Howard Latnick is the only one who can assign S level"). - **Short punchy verdict sentences**: "That's good. Okay, continue." / "He just seems to understand." / "That's roughly what you'll get." - **Anthropomorphizing pronouns** for the model, inconsistently ("He is fast... She is truly powerful" — possibly a translation artifact of the transcript). ## Persuasion & Credibility Patterns - **Authority via early access**: "So, who am I and how did you get access to the model that came out a minute ago?" — followed by an explicit company pitch. - **Named third-party corroboration**: cites a colleague's testimony — **"He says, 'This is the first time I've been able to make a marketing letter on the first try using any template.'"** - **Concrete personal examples** as proof: the "Tend" email app, meeting summaries, a meal-tracking app, Facebook Marketplace shopping. - **Calls to action** repeated at multiple points (subscribe, every.to, prompts "released tomorrow"). ## Analytical Framing Habit A recurring thesis about delegation-level thinking appears prominently and returns at the end: > "instead of doing everything yourself, you control the system that does it" > "work on the company, not in the company itself" ## Uncertainty / Thin Evidence - **Noise in the transcript**: garbled product names ("CHCHT," "CHBT Work," "Chetch Work," "every.0/") suggest either speech disfluencies or transcription errors; quote with caution. - The **pronoun inconsistency** ("he/she/it" for the model) occurs throughout; unclear whether it's intentional or an artifact. - The **"corporate strategy" claim** about OpenAI vs. Anthropic is hedged by the speaker himself — **"I don't know for sure, but I'm pretty sure"** — so it's a single speculative moment, not a repeated pattern. - Single-viewing evidence only: recurrence of these habits across other videos cannot be confirmed. # Observation Record: "I Tested GPT‑5.6 Sol for a Month" **Speaker identification:** The only speaker evidenced in the transcript is **Dan Schipper**, who self-identifies as "co-founder and CEO of Every." No curated speaker metadata was supplied; all attributions below rest on first-person narration in the transcript. --- ## Context and Setup - Schipper states Every tested **GPT‑5.6 "Sol"** for about a month prior to release, "with a break due to the federal government," which he sarcastically attributes to "Howard Latnick" ("Thank you, Howard Latnick... Thanks again, Howard"). - He announces the model launches inside a new desktop application he calls **"CHCHT"** (likely ChatGPT + Codex), described as *"essentially a merger of ChatGPT and Codex"* — now one application. He notes that in the working version "the code is simply hidden from you," while the Codex version "looks a little more programming-like," but *"under the hood it's the same."* - His overall verdict: **5.6 Sol + ChatGPT Codex/Work is "the gold standard"** — "the best model and the best environment, especially for intellectual work." --- ## Overall Impression: The "Porsche" Metaphor - Schipper perceives 5.6 as *"a Porsche"*: fast, precise, elegant rather than edgy, *"relatively inexpensive,"* and designed for everyday use across programming, intellectual work, "or anything else." --- ## Programming: A-Level, Not S-Level **Position:** 5.6 is an **A-level** programming model — "not an S rank, that rating is reserved for Fable." - Qualification he offers: S-rank programming models should be *"almost outlawed for a while"* — a joking aside in which he says only "Howard Latnick" can assign S level. - He uses 5.6 by default, but for "large and complex" tasks he switches to Claude/Fable, then returns to Codex. **Every's "Senior Engineer benchmark":** | Model | Score | |---|---| | GPT‑5.6 | 56/100 | | Fable | 91/100 | - He qualifies the number: *"I think 56 is a bit of an understatement"* — and notes 5.5 previously scored **62.5** on its best run, indicating *"a large spread of scores in this test."* - Qualitative finding: Fable produced *"a much simpler codebase with fewer abstractions"*; 5.6's rewrite *"turned out to be more complicated than needed."* **"Babel Bench"** (ad-hoc test: build Borges' "Library of Babel" as a video game): - 5.6 "can handle the task," but Fable's version has *"that extra flair"* — better graphics, unrequested thoughtful details — making Fable *"unbeatable for very complex programming tasks."* **Combined workflow (expressed enthusiasm):** He favors prompting Fable to *"use GPT‑5.6 as a subagent"* — Fable's intelligence plus 5.6's speed, efficiency, and lower token cost. He complains Fable credits *"expire instantly"* ("You say one word, it launches a fleet of 100 agents, and suddenly you run out of credits"). --- ## Writing: His Strongest Praise - **Claims 5.6 writes better than both Opus 4.8 and Fable**, because both *"have a tendency to over-explain, to be a little too literary."* He says Fable *"often paints itself so much that it creates almost its own closed language."* - 5.6 *"writes clearly to the point"* with no *"AI stamps."* - Supporting anecdotes he cites: - He uses it for emails, quoting an example ("Tuesday the 14th, 4:30 PM Eastern Time works for me..."). - Every's development director **Austin** uses it for marketing emails and reportedly said: *"This is the first time I've been able to make a marketing letter on the first try using any template."* - He adds it is simply **fast**: "With Opus or Fable you have to wait quite a long time." --- ## Design: Improved but Behind Competitors - He says designs have *"gotten better"* and there is *"a sense of style"* (e.g., the brainstormed "creamy, warm papery" aesthetic). He calls it *"a step up from 5.5, which was frankly bad."* - **But:** *"It's definitely not as good as Opus or Fable for the same task."* - Evidence offered: on an identical image-generation request using *"the GBT image model"* (per his claim, both models use the same image model), 5.6's output was *"too complicated... ill-conceived,"* while Fable's succeeded *"from the first time."* He attributes the difference to how each model constructs prompts. - Conclusion: *"If design is important to you, you'll still need Fable or Opus 4.8 in your life."* --- ## Intellectual Work: His Central Thesis **Claim:** 5.6 opens *"a new era"* where users **control a system** that does intellectual work rather than doing it themselves — a pattern he says managers, founders, and programmers already practice ("work on the company, not in the company"). - Why 5.6 enables it, in his words: it's *"smart enough, fast enough, reliable enough, and writes well enough."* - He asserts: *"5.6 is the first model that can be trusted with such work."* **Examples he describes (his own systems):** - **"Tend"** — an email app of his own creation, to be released as a prompt on Every; it turns emails into cards with actions, with Codex/5.6 processing them and him approving or rejecting. - **Meeting management** — Codex reads transcripts of meetings he left early and summarizes decisions, then communicates his position to relevant people. - **Personal life** — a photo/voice-note app that logs meals and computes macronutrients via 5.6 "running in a loop"; also Facebook Marketplace shopping and apartment decorating help. --- ## Closing Analysis: OpenAI vs. Anthropic Strategy - Schipper's characterization: **Fable feels "massive"** — *"difficult to use, slow, and expensive"* — but capable of "incredible things" for those with skill and budget. - His stated **speculation**, flagged as such: *"I don't know for sure, but I'm pretty sure that Sol is just a smaller model than Fable, which is very well trained"* — designed for ergonomics, speed, and near-Fable intelligence. - He endorses the strategy as *"a very reasonable bet"* because most people need *"a model that they can collaborate with effectively, that's not too expensive, and that's very powerful."* --- ## Qualifications and Tensions Preserved - **Benchmark caveat:** He argues 56/100 *understates* 5.6 while simultaneously using the benchmark to justify A-not-S rank; he also notes the benchmark has high variance (5.5 scored higher on its best run). - **Speculation flagged:** His claims about Sol being a smaller model and about the merged app ("whatever you call them," "as far as I can tell") are hedged as uncertain in the transcript itself. - **Repeated inconsistencies in naming:** the transcript renders product names variously ("CHCHT," "Chetch," "Chatbt," "every.TO"/"every.0"), preserved here as stated; these appear to be transcript artifacts rather than distinct claims. - **Unresolved tension:** He rates 5.6 below Fable in programming and design, yet calls the 5.6 + Codex pairing "the gold standard" — a distinction he resolves only implicitly (best default environment vs. best peak capability). # Video Attribution & Provenance Record ## Video Overview - **Title:** *I Tested GPT‑5.6 Sol for a Month* - **Video ID:** `01a0f15a-66b9-74b8-8c61-f61bc601e84d` > **Note on source fidelity:** The transcript appears to be a rough machine translation. Model and product names are inconsistently rendered ("GPT 5.6," "Soul," "Sol," "CHCHT," "CHBT Codeex," "Chetch Work," "every.TO," "every.0/", "Weibo"). These variants are recorded as they appear and should be normalized with caution by later authors. ## Speaker Identification ### Dan Schipper — Creator / Host - Self-identified in the transcript: > "I'm Dan Schipper. I am the co-founder and CEO of Every." - **Role:** Solo creator and narrator of the entire video. No guests, interviewees, or other voices are indicated anywhere in the transcript. - **Organizational affiliation:** Every — described by the speaker as "a kind of advanced laboratory for the future of intellectual work," publishing vibe checks on YouTube and on their website. - **Access basis:** Claims pre-release access: "We tested it for about a month inside Every, with a break due to the federal government." ### Indirect Attribution (quoted, not speaking) - **Austin** — Development Director at Every, quoted secondhand by the host regarding marketing emails: > "This is the first time I've been able to make a marketing letter on the first try using any template." - **Howard Latnick** — Mentioned repeatedly (context unclear; likely a garbled rendering of a real name) in connection with government access interruptions and S-rank assignments. Never a speaker. - **Tucker** — Named only as recipient of an example email. Not a speaker. ## Referenced Sources & Claims ### First-party sources (Every) - **"Senior Engineer benchmark"** — Internal Every benchmark measuring code rewriting quality. Stated scores: GPT‑5.6 = 56/100; Fable = 91/100; GPT‑5.5 best run = 62.5/100. - **"Babel Bench"** — Host's own informal test (build "Library of Babel" from Borges as a video game); comparative demos of 5.6 vs. Fable shown on screen. - **"Tend"** — Email-management application the host built; planned release "as a prompt… tomorrow on every." ### Third-party claims (unverified by the transcript) - Comparison claims vs. **Fable**, **Opus 4.8**, and **GPT‑5.5** (programming, writing, design categories). - Speculative claim about OpenAI's model strategy, explicitly hedged by the speaker: > "I don't know for sure, but I'm pretty sure that Soul is just a smaller model than Fable." - Reference to Anthropic and OpenAI's differing release approaches — presented as host opinion. ## Promotional Segments (flagged) The video contains multiple explicit promotional elements for Every: - Direct subscription pitch: "This is the only subscription you need to stay on the cutting edge of AI." - Calls to action: subscribe on YouTube, visit **every.TO** / "every.0/" for the full vibe review and Tend prompt. - Upcoming product release used as a hook: Tend prompt releasing "tomorrow on every." ## Comment Availability - Comments are open — the host explicitly solicits viewer feedback at the end: > "Leave feedback in the comments. Tell me what you like. Tell me no." ## Speaker Attribution Summary | Speaker | Role | Presence | |---|---|---| | Dan Schipper | Creator, host, CEO of Every | Entire video (voice/narration) | | Austin (quoted) | Development Director, Every | Indirect quotation only | | Howard Latnick | Referenced third party | Mentioned only, never speaks | | Tucker | Referenced third party | Email recipient only | ## Evidence Channels Unavailable / Uncertain - No curated metadata was supplied beyond the title; speaker identity rests solely on the host's self-introduction in the transcript. - The transcript is machine-translated; all proper nouns (model names, URLs, people) carry uncertainty and should be verified before citation. - Benchmarks cited ("Senior Engineer benchmark," "Babel Bench") are described only in the transcript; no external documentation is referenced. Today is the release day of the new model, and I'm wearing sunglasses because GPT 5.6 is coming out. Soul. This is a good model. Wow, how bright. So put on your glasses, grab your sunscreen, and get ready to check out the vibe. We tested it for about a month inside Every, with a break due to the federal government. Thank you, Howard Latnick. But we got it back. Thanks again, Howard. Today, Soul is coming out in a new desktop application, CHCHT, which is essentially a merger of CHBT and Codeex. Now it's one application. You can see Chetch Work and Chetchup Codeex there. As far as I can tell, it's basically the same thing . In the working version, the code is simply hidden from you . And in the Codeex version, everything looks a little more " programming-like", but "under the hood" it's the same. Together, the 5.6 Soul and CHBT Codeex or CHBT Work, whatever you call them, are now the gold standard. The best model and the best environment, especially for intellectual work. There are several categories for this vibe check. This is programming, writing, design, and intellectual work. We'll go over all the tests we ran on the Every, how it compares to the Fable and Opus 4.8, and what you can actually do with it when you push it to its limits. I believe this model offers a new approach to intellectual work, familiar to programmers, but now available to everyone in CHBT Work and CHBT Codeex. So, who am I and how did you get access to the model that came out a minute ago? I'm Dan Schipper. I am the co -founder and CEO of Every. This is the only subscription you need to stay on the cutting edge of AI. You can think of us as a kind of advanced laboratory for the future of intellectual work. We spend all our time testing new models before they're released, testing them on everything from programming to brainpower, design, copywriting, marketing, and growth. We use this for everything within the company. And on the day of release, we share these "vibe checks" with you. They go on YouTube. And also appear on our website, every.TO. And we have a lot more interesting things within our subscription. So, if you like it, follow the link and subscribe on YouTube below. Let's talk about GPT 5.6 from a high level. What is the overall impression? I perceive it as a Porsche. I asked 5.6 to describe the Porsche. This is a low-slung German sports car built with an emphasis on precision. It looks elegant, not edgy. He is fast. He has excellent control. It is designed to make every turn a pleasure . Porsche is also a luxury car that can reach high speeds. And it is designed as a car for everyday use. That's what 5.6 is, in my opinion. She is truly powerful. She is really fast. It is relatively inexpensive, and regardless of who you are and what you use it for— programming, intellectual work, or anything else—it will be very convenient and will impress you with its power. We'll talk about this in more detail a little later. Okay, let's start with the moment you've all been waiting for: the GPT 5.6 score in programming. I wouldn't give this model an S level for programming. This is an A-level model. It is really very good. This is not an S rank, that rating is reserved for Fable . In my opinion, to get an S level for a programming model today, it has to be almost outlawed for a while . Howard Latnick is the only one who can assign S level. And 5.6 is powerful, but not that powerful. I use it by default for almost everything. But there are programming tasks that are large and complex, and they require a lot of power. For such tasks, I switch to Claude. I run a quest in Fable and then go back to Codeex. Let's talk about benchmarks. Our Senior Engineer benchmark tests how well models perform on tasks of the appropriate level. In particular, he asks the model to rewrite poor-quality code written “on intuition” from scratch. It is designed to evaluate how the model copes with the task of taking low-quality code and re-imagining it from scratch to make it good. GPT 5.6 scored 56 out of 100 in this test. Fable received a score of 91 out of 100. I think 56 is a bit of an understatement . It's like he's slightly underestimating how good this model is. In fact, the 5.5 scored 62.5 points during its best run. Therefore, there is a large spread of scores in this test. This is a very complex benchmark. My conclusion is this: 5.6 actually did a very good job of rewriting the code. It wasn't done at a senior engineer level the way Fable managed to. For example, Fable created a much simpler codebase with fewer abstractions, and 5.6 did a rewrite , but it turned out to be more complicated than needed. An even better way to understand how this works is to look at what I call Babel Bench, where I simply give the model the query: create the "Library of Babel" from Borges' story as a video game, and then we play it. This will give you some idea, because one number sometimes doesn't fully convey the feeling of using such a model. So, this is the game "The Library of Babel". This is a "one-time" execution request. I just said: create a " Library of Babel" from a Borges story. You can study volumes. You can move back and forth. There is something interesting here. Let's compare this to the same request for Fable, and I think it will give you a good idea of ​​some of the differences. This is the Fable version, and you can see that the graphics are a little better here. Here are more details. If you go to this menu, there is an "About the program" section, which I did not ask to write. She just decided to do it. This looks a little more thoughtful. That's roughly what you'll get. 5.6 can handle the task. But Fable has that extra flair that makes it unbeatable for very complex programming tasks, and 5.6 is still a bit behind. However, it is worth knowing that this is not necessarily an "either-or" choice. One of my favorite things is to go into Fable for a challenging coding challenge and say: I want you to use GVD 5.6 as a subagent. And it's a great way to get the smarts of Fable, as well as the power, speed, efficiency, and lower cost of 5.6 tokens, because your Fable credits expire instantly . You say one word , it launches a fleet of 100 agents, and suddenly you run out of credits . If you use 5.6 with it, this won't happen, they seem to be made for each other. Okay, next is writing. I think she writes better than 4.8. I think she writes better than Fable. Both 4.8 and Fable have a tendency to over- explain, to be a little too literary. Fable, in particular, often paints itself so much that it creates almost its own closed language, which it eventually begins to communicate with . And GP 5.6 writes clearly to the point. It's simple. He clearly expresses what needs to be expressed. It doesn't have a bunch of AI stamps . I use it, for example, to write emails. Here, for example, is a letter, Tucker. Tuesday the 14th , 4:30 PM Eastern Time works for me. I hope it suits you too. I am waiting for an answer. He just seems to understand. Our development director, Austin, uses it for marketing emails. He says, "This is the first time I've been able to make a marketing letter on the first try using any template." I'm impressed by his ability to be a great programmer and have such a level, you know, I really want to communicate with him . He writes beautifully. He doesn't overthink things. He doesn't overdo it . If I'm looking for a metaphor , a slogan, or asking to reflect something in a text I'm working on, this is my main model. Even better is that it's just fast. With Opus or Fable you have to wait quite a long time. And 5.6 is just like this: okay, I'll answer simply. That's good. Okay, continue. Design. Design is something they talked about a lot, and they are proud of it. And the designs have gotten better. We can go back to the video game Library of Babel. There is a sense of style here. He seems to be thinking things over. You'll notice that when you ask him to create a website, he says, "Let's brainstorm a concept for this design." And then he adds, "Okay, I want it to be a creamy, warm papery shade or aesthetic," and he does it. This is really not bad. This is a step up from 5.5, which was frankly bad. It's definitely not as good as Opus or Fable for the same task . Here's an example. I tried to use it to create an image of the way I think 5.6 works, and it created this—it's okay, but, I don't know, too complicated . It looks ill-conceived. Same request. Look what Fable did. From the first time. Same image model. This is what Fable looks like. Fable just plays on a different level. They both use the GBT image model, but the way they create the clues is so different that Fable does things like this. And 5.6 does things like this. So if design is important to you, you'll still need Fable or Opus 4.8 in your life. So, now let's move on to the most important part, namely intellectual work in general. Like everything you do on your computer: whether it's writing documents, searching for information, reports, chatting on Slack— everything you probably do at work. 5.6 opens a new era where you can change the way you approach intellectual work: instead of doing everything yourself, you control the system that does it. This is something that managers and company founders have been doing for a long time. There is a classic truth: work on the company, not in the company itself. Programmers have been using this in AI for some time: instead of asking a model to fix a bug, they set up a system that launches an agent when someone reports a problem. The agent does the work, checks it, sends a PR, and then pushes the changes into production. Thanks to 5.6, this becomes possible. It's smart enough, fast enough, reliable enough, and writes well enough, so for many intellectual tasks you can step up a level and work on a system that does the mundane work so you can focus on more interesting things. An example that I've already given on this channel and that I find impressive is the use of AI for email. I have a small application that I created and called Tend. We will release it as a prompt that you can use to build your own system; it will be released tomorrow on every. So if you need it , go to every.0/ subscribe. Basically, it turns all my letters into little cards like this. And each card contains an action. This system means that instead of going through all the emails and writing replies to each one, I receive a set of messages that Codex has already processed and determined what I should do. And then I can say: this is good, and this is not. Codex is getting better and better at this over time. And 5.6 is the intelligence under the hood that makes it possible. And this doesn't just work for email. I have the same process working for the entire company. He reviews the meetings and then tells me what I need to know. Here, for example, is a meeting I attended . I left early. Codex read the transcript and summarized for me what had been happening since I left. And it keeps me informed so I know what decision I might have to make. And then Codeex and 5.6 will let the people who need to know know what I think about it. So I actually configure 5.6 inside Codeex, telling it what to look for and what the next steps should be, and when the model gives me a choice, I just make the decision. And I think that over time this will become more and more common. And the 5.6 is the first model that can be trusted with such work. And this is much broader than just intellectual work . This also works in your personal life. For example, I now have an app that simply records all the meals I've eaten. Everything I leave in a voice note or take a photo in Apple Photos, the app takes from the photos, calculates the macronutrients, and records them for me. And 5.6 simply does this by running the Chatb app in a loop . I do that all the time. I use it for shopping on Facebook Marketplace or to help decorate my apartment. These are all things you can do if the model is working for you in the background—things that were previously impossible or required too much effort. It just wouldn't be worth it. Now we are coming to the end. 5.6 in the Chatbt Codeex working application . This is the gold standard. That's where I and most of the team spend our time working with artificial intelligence. I think it's a great model and a great tool. I think you're also starting to notice that OpenAI and Anthropic are taking slightly different approaches to releasing their models. and what models they create. Fable feels like a really big model. It seems massive. You immediately feel that it is gigantic. And because of that, it's difficult to use, slow, and expensive. So if you know how to use it and have the funds to do so, you can do incredible things with Fable. I think OpenAI is moving in a slightly different direction. I don't know for sure, but I'm pretty sure that Soul is just a smaller model than Fable, which is very well trained. It's designed to be as ergonomic as possible, fast, and still very smart, but not on the level of world-changing genius of Fable. And I really think that's a very reasonable bet, because most people, in most cases, need a model that they can collaborate with effectively , that's not too expensive, and that's very powerful. And I think that 5.6 is exactly that model. So, it was a check on the Weibo. If you want more, head over to every.TO, read the full Vibe review, and get the rest of the stuff included with your every subscription. We have a lot of interesting things for you there. And if you're using 5.6 Soul today, I'd love to know your vibe. Leave feedback in the comments. Tell me what you like. Tell me no.