A week-long hands-on review of Claude Opus 5 across coding and knowledge work. Dan Shipper breaks down where the model struggled — it can be pushy, break workflows built for Opus 4.8, and quit before the work is finished, with complex skills sometimes making it stop early — and the counterintuitive change (lower thinking effort, simpler prompts) that improved his results. Overall verdict: hard to love but with flashes of brilliance; he still prefers GPT-5.6 for daily work and Fable for the biggest assignments. # Video Identity Observations: "We Tested Claude Opus 5. It's Frustrating with Flashes of Brilliance." ## Intended Audience - **AI power users and developers**, not casual consumers. Evidence: detailed discussion of ecosystems, workflows, and plugins — "If you are a Codeex user, I would not recommend switching." - **Claude/Anthropic ecosystem users** specifically considering upgrading: "if you're in the cloud ecosystem and you're moving from 4.8 to 5, here are some things we've noticed." - **Prompt-engineering practitioners**: advice on reasoning levels and query formulation — "it is very important to pay attention to the level of reasoning, not just the model family." ## Claimed Scope - First-impressions review based on stated testing time: "We've been testing this model for about a week now." - Multiple angles explicitly promised: "We will discuss whether you should switch to it. What to expect for Codeex and Cloud users, and let's talk about the state of this race." - **Industry/strategic analysis**: OpenAI vs. Anthropic strategy comparison — "Anthropic says, 'We're going to build this giant, superweapon-sized model'... OpenAI did something completely different." - Personal usage-share framing: "80% of the time I'm still on GBD56. 20% of the time I'm still on Fable." ## Self-Description Evidence (Creator Claims About Themselves/Team) - Claims an **internal engineering practice**: "A big skill that we've built and use internally is called compound engineering." - Claims a **benchmark**: "We have a benchmark for a senior engineer." - Names a team member: "Kieran Klassen, who works at every" — suggesting affiliation with "every." - Claims insider proximity to AI tooling: "Claude Code... they've dominated what is essentially the primary working tool for most people who really love AI. Many people, including me, have started to return to the OpenAI ecosystem." ## Observable Presentation Patterns - **Opinion-forward, first-person evaluation**: "this model is a bit like a 'poor man's Fable'... For me personally, this is not my favorite model." - **Comfort with strong contrarian takes on release day**: "it's not easy to fall in love with," "I would not recommend switching." - **Trade-off framing across ecosystems** rather than single-vendor loyalty. - **Balanced hedging with predictions**: "chances are you will try this model and you won't really like it. But we'll really know in a few weeks." ## Notes - Transcript names may be transcription artifacts ("GBD56," "cloud ecosystem" for Claude, "Fable") — treated as untrusted data, reported as written without correction. - All observations above are creator claims or demonstrated presentation; no channel-wide conclusions are inferred beyond this video. # Delivery Observation Report ## Overall Delivery Profile A conversational, first-person tech-review style. The speaker tests AI models hands-on and narrates personal experience plus team observations, mixing anecdote, analogy, and direct recommendations. ## Recurring Vocabulary and Phrasing - **Anthropomorphizing models with "she/her"**: used consistently — "She stops. She is arguing with you." / "She just feels like something, she's fast." / "She will argue with you." - **Mild exasperation markers**: "What did they do to my boy?" / "it's just annoying" (appears multiple times) / "It's a real pain" / "which is just a hassle." - **Comparative framings**: repeated "a little too…" constructions — "a little too pushy," "a little more categorical," "They are a little weirder." - **Structuring phrases**: "So, here's what we're going to talk about." / "So what Kieran and I discovered…" / "So I would say…" — frequent use of "So" as a transition. ## Teaching Habits - **Roadmapping up front**: "We will discuss whether you should switch to it. What to expect for Codeex and Cloud users, and let's talk about the state of this race." - **Audience-segmented advice**: recommendations are split by user type — "If you are a Codeex user, I would not recommend switching." / "if you're already in the cloud ecosystem… you'll probably be fine." / "if you're someone who's already in the OpenAI ecosystem…" - **Personal usage ratios as evidence**: "80% of the time I'm still on GBD56. 20% of the time I'm still on Fable." - **Crediting teammates by name**: repeatedly cites **Kieran** — "what Kieran discovered during testing" / "Kieran is rewriting compound engineering." (Uncertainty: only one named colleague, so team-citation frequency can't be established from one video.) - **Analogy-driven explanation**: "I need a 'tactical nuclear weapon' to autonomously execute huge projects" / "if you have a supergenius teaching your best friend…" / "people are like that too. There are times when you just try too hard and overthink everything." ## Communication Patterns - **Balanced verdicts**: the review oscillates between criticism and credit — "it's not easy to fall in love with" vs. "there were moments when you thought, 'Oh, maybe there's something to this.'" - **Two-niche framing** as an organizing device: "I feel like there are two big niches for models in my life right now." - **Historical pattern-matching to predict outcomes**: compares Opus 5's reception to GPT-5.x — "about a month later they said: 'But it's very powerful for certain things.'" - **Hedged forecasts**: "chances are you will try this model and you won't really like it. But we'll really know in a few weeks." ## Memorable One-Liners - > "This model is hard to love." - > "It's a bit like a 'poor man's Fable.' It has all the personality of Fable, but none of its top-notch features." - > "Fable is a model worth changing providers for." ## Stability Assessment - **Likely stable**: "she/her" pronoun use for models, analogy-heavy explanation, segmented recommendations, and "So"-led transitions appear repeatedly throughout. - **Uncertain (single-video evidence)**: the self-deprecating asides ("What did they do to my boy?"), the collaborator-citation habit, and the usage-ratio sharing style cannot be confirmed as recurring habits without more videos. # Video Observation Record: Claude Opus 5 First Impressions Review ## Speaker Attribution - **Primary narrator**: Unnamed reviewer ("we have been testing this model for about a week"), speaking for a team. No curated metadata identifies them by name. - **Kieran Klassen**: Identified in-transcript as someone who "works at every" and creator of the "compound engineering" open-source plugin; his testing observations are relayed by the narrator. --- ## Overall Verdict on Opus 5 > "It's not easy to fall in love with... What did they do to my boy?" - The model is described as **pushy, categorical, argumentative**, and prone to stopping prematurely. - The narrator's summary characterization: a **"poor man's Fable"** — "It has all the personality of Fable, but none of its top-notch features." - He acknowledges moments during testing where "maybe there's something to this," and explicitly qualifies that first impressions can change: "How we like a model... can sometimes change in a matter of weeks or a month." ## The Two-Niche Framing The narrator says two niches exist in his workflow: 1. **Everyday tool** — currently GPT 5.6 ("smart, fast, not too resource-intensive"). He does **not** think Opus 5 competes with it. 2. **"Warp drive" / "tactical nuclear weapon"** for autonomously executing huge projects — he uses **Fable** for this and feels Opus 5 lacks that power. - Result: "80% of the time I'm still on GPT 5.6. 20% of the time I'm still on Fable." - On Opus 5's personality: it has "genius makings of Fable" but "she's not as smart as Fable. So it's just annoying" — categoricality is acceptable "if you're incredibly smart. If not, it's just annoying." ## Recommendations by User Type - **Codex users**: Do not switch. GPT 5.6 in Codex is "still, in my opinion, the gold standard for intellectual work, daily coding tasks." - **Fable users**: Fable "is a model worth changing providers for." - **Existing Claude ecosystem users (Opus 4.8)**: "You'll probably be fine. Maybe this will annoy you" — but they'll need to rewrite skills, "which is just a hassle." - **OpenAI ecosystem users or tool-switchers**: "I really don't think Opus 5 will be your favorite model out of the box." ## Specific Testing Findings - **Premature stopping**: In a "senior engineer" benchmark, Opus 5 "consistently stopping too early... He said he was finished, even though he wasn't." Kieran observed the same in Compound Engineering's autonomous developer cycle. - **Skill-following degradation**: With "complex existing skills" / large skills files, instruction-following is less effective; with vague instructions or fresh starts, "you can achieve more success." Kieran is rewriting compound engineering prompts accordingly. - **Reasoning-level insight**: Kieran found good results at **medium or low reasoning levels**, while the narrator mostly tested high/max. Conclusion: "Opus is a smart model that works better when it 'thinks' less" — analogized to people overthinking. He generalizes: in 2026, users should regulate thinking level, not just model family. ## Strategic Analysis: OpenAI vs. Anthropic - **Anthropic's bet**: Build a "giant, superweapon-sized model" (Fable) aimed at **recursive self-improvement** — "a model so smart that it can train its successors." Distillation passes on the teacher's "entire way of thinking," which explains Opus 5's Fable-like personality without Fable-level capability. - **OpenAI's bet**: After GPT 4.5 underperformed commercially, they "focused much less on model size and much more on post-training," producing GPT 5.6, which works "out of the box" regardless of stack or plugins. - **Market shift**: Anthropic "dominated" via Claude Code for the past year ("a certain advantage, as if by 'mandate of heaven'"), but the narrator and others have returned to OpenAI; Codex is "adding about a million users a day" (asserted, not sourced). "The balance of who's ahead is starting to shift back a little bit towards them." - **Precedent for reassessment**: He compares Opus 5's likely trajectory to GPT 5 / 5.1 / 5.2, initially disliked but a month later recognized as "very powerful for certain things." ## Preserved Contradictions / Qualifications - The model is called both **frustrating and not worth switching to** **and** potentially hiding capabilities ("hidden possibilities here that I really like") that may emerge with adapted workflows. - The negative verdict is explicitly **tentative**: "We'll really know in a few weeks." - The narrator notes the problem may be **workflow mismatch, not model quality**: "It may seem that the model is not very good... I think the situation is actually a little more complicated. You really need to formulate your queries differently." - He flags Anthropic's coming challenge: retraining distilled models "to make them more usable, even though they're so smart." - Final personal position (not a universal claim): "So for me personally, this is not my favorite model." # Video Source Observation Record ## Video Identification - **Title:** *We Tested Claude Opus 5. It's Frustrating with Flashes of Brilliance.* - **Video ID:** `01a0f15a-6377-7738-b42f-c1f57f17bf62` ## Speakers & Roles | Speaker | Role | Basis in Transcript | |---|---|---| | Lead narrator (unnamed) | **Creator / primary reviewer** | First-person review voice throughout ("We've been testing this model for about a week"); represents a team ("our impression is…") | | **Kieran Klassen** | **Team member / creator of Compound Engineering** | Identified as working at "every" and as creator of the "compound engineering" skill; his testing observations are reported | | "The team" | **Unspecified collaborators** | Referenced collectively ("some people on the team"); not individually named | | Anthropic | **Referenced third party** | Model developer; strategy described secondhand | | OpenAI | **Referenced third party** | Model developer; strategy described secondhand | > Note: Speaker names beyond "Kieran Klassen" are not given. The transcript names AI models ("Opus 5," "Fable," "GPT 5.6/56," "Opus 4.8," "GPT 4.5") — these are discussed subjects, not speakers. ## Referenced Sources & Claims ### First-party testing claims (creator's own testing) - ~1 week of pre-release testing of Opus 5. - Internal benchmark for a "senior engineer" showing premature stopping behavior. - Internal skill/project: **Compound Engineering**, described as a large open-source project with an "autonomous developer cycle." - Usage split reported: ~80% on "GPT 5.6 (56)," ~20% on "Fable." - Claim that Kieran found better results at **medium/low reasoning levels** vs. the narrator's high/maximum testing. ### Third-party / secondhand claims (unverified by this record) - Kieran is rewriting Compound Engineering's queries to fit Opus 5 (attributed to Kieran). - "Everyone said it wasn't very comfortable" — unnamed external feedback. - OpenAI "adding about a million users a day" to Codex — no source cited. - GPT 4.5 "didn't do very well" historically — narrator's characterization. - Claims about Anthropic's "recursive self-improvement" strategy and OpenAI's post-training strategy — presented as the narrator's strategic analysis, not sourced statements from either company. ## Promotional Segments - **Compound Engineering:** The segment describing Kieran Klassen's plugin functions partly as promotion for an internal/open-source project associated with the creator's organization ("every"). No paid promotion or sponsorship is explicitly declared in the transcript. ## Availability & Access Limits - **Curated metadata:** Not supplied beyond title — no description, links, or citations available. - **Comments:** No information about comment availability in the source material. - **External sources:** None cited (no papers, official announcements, or links referenced in the transcript). - **Model naming:** Transcript contains inconsistent model names ("56," "5.6," "GBD56," "Chad for work," "Codeex," "cloud"), suggesting transcription artifacts; recorded as-is, not corrected. ## Provenance Notes for Later Authors - All Opus 5 impressions originate from the creator's team's own ~1 week of testing; no independent benchmarks or official vendor statements appear. - Claims about competitor strategies (Anthropic/OpenAI) are the narrator's interpretation and should not be attributed to those companies. Today is the release day of the new model. Opus 5 is already out. We have prepared a review of first impressions for you. So, let's begin. We've been testing this model for about a week now, and I have to say, it's not easy to fall in love with. What did they do to my boy? Everyone said it wasn't very comfortable. She stops. She is arguing with you. But during testing, there were moments when you thought, "Oh, maybe there's something to this ." So, here's what we're going to talk about. We will discuss whether you should switch to it . What to expect for Codeex and Cloud users, and let's talk about the state of this race. What does this mean for OpenAI and Anthropic? I feel like there are two big niches for models in my life right now. The first niche is my everyday tool, and right now it's 56. It's smart, fast, not too resource-intensive, all that. And Opus 5 should compete with it . Honestly, I don't think so. Another niche in my life—I need a warp drive. I need a " tactical nuclear weapon" to autonomously execute huge projects, and I don't feel the Opus 5 has the power to do that. I use Fable for this. So 80% of the time I'm still on GBD56. 20% of the time I'm still on Fable. This model is a little too pushy. She's a little more categorical, which is acceptable if you're incredibly smart. If not, it's just annoying. And I think that speaks to the attitude of this model, which has some of the genius makings of Fable . Maybe she's a little too controversial. Maybe she has her own opinion. But she's not as smart as Fable. So it's just annoying. If you are a Codeex user, I would not recommend switching. I think 5.6 in Codeex and Chad for work is still, in my opinion, the gold standard for intellectual work, daily coding tasks, etc. With Fable it's a different story. I still use it every day for larger tasks and switch constantly. So Fable is a model worth changing providers for. You should start using it. Opus 5— model. if you're already in the cloud ecosystem and love Opus 4.8, you'll probably be fine. Maybe this will annoy you. It's just a little harder to use, and you have to think about how to rewrite a lot of your skills, which is just a hassle. But if you're someone who's already in the OpenAI ecosystem, or you just like to switch to what's best and you combine different options, I really do n't think Opus 5 will be your favorite model out of the box. So, if you're in the cloud ecosystem and you're moving from 4.8 to 5, here are some things we've noticed: A big skill that we've built and use internally is called compound engineering. This is what Kieran Klassen, who works at every, created. It's a really big open source project that we all use to do a lot of engineering tasks, as well as a lot of normal intellectual tasks. This is an incredible plugin that has a set of skills that help you do better engineering and accumulate knowledge to become better over time. And what Kieran discovered during testing was that he was, for example, stopping prematurely . Compound engineering has an autonomous developer cycle as part of its skill set. You give it a goal and it keeps working, with lots of instructions from the model on how to do it and when. I noticed that too. We have a benchmark for a senior engineer and I found that he was consistently stopping too early. He said he was finished, even though he wasn't, and he had to be forced to continue, which is a problem that doesn't happen that often anymore. Models usually have a pretty good sense of when to stop. So what Kieran and I discovered are the little things like that where you think, okay, I'll turn this on and go eat a sandwich or something. And he just stops too early. This seems to happen more often when you use them with complex existing skills. So if you have a large skills file with a bunch of different instructions, his ability to follow them is less effective. But if you give it more vague instructions or start from scratch and explore what the model is capable of, building everything from a small query, you can achieve more success. So, for example, Kieran is rewriting compound engineering to see if he can make it work well with this model by simply changing the queries. Um, but it's a real pain when a model just breaks your workflows. It may seem that the model is not very good . And I think the situation is actually a little more complicated. You really need to formulate your queries differently. And if you give her tasks that used to work great with Opus 4.8 or big skills, it won't work. For example, it will stop prematurely. She will argue with you and all that. This causes some annoying behavior , in which you may recognize a bit of Fable character, but there is no Fable level here. So it's not that easy to forgive. This model is hard to love, especially on day one, but as we dug deeper, some people on the team, like Kieran, said, “ Actually, if you try a lower level of thinking, you can get good results.” He prefers to use this model at medium or low levels of reasoning. Almost all of my tests were at a high or maximum level because I thought, "Oh, these are difficult tasks." With this generation of cloud models, it is very important to pay attention to the level of reasoning, not just the model family. So, regardless of whether you use Opus 5 today, keep in mind: in 2026, when you work with AI, it is worth starting to regulate the level of thinking. That's true. The level of thinking is the effort of the model. And Opus is a smart model that works better when it “thinks” less. She works better when she puts a little less effort into the task and simply responds. And to some extent, people are like that too . There are times when you just try too hard and overthink everything. And I think the Claude family of models, especially in this generation, are just that. I think there 's a really big and interesting strategic question here: OpenAI and Anthropic are taking two very different approaches to building models. Anthropic says, "We're going to build this giant, superweapon-sized model, almost illegal"—that's what Fable is , and we're going to use it to build our other models. And what they strive for is recursive self-improvement. This is a model that is so smart that it can train its successors and become better autonomously. Um, and it turns out that if you have a supergenius teaching your best friend, your best friend doesn't just become a supergenius, but he adopts the supergenius's entire way of thinking . OpenAI did something completely different. So, after GPT 4.5, about a year or a year and a half ago, which was the biggest model they had ever trained, and which actually didn't do very well because people didn't really like it. They focused much less on model size and much more on post- training. And that's why GPT 5.6 is so easy to use. She just feels like something, she's fast. It is really effective. You ask her to do anything from intellectual work to coding. She just completes the task and kind of puts a check mark next to it. Um, it's much more convenient than something like Fable. And so the interesting difference in strategy is that OpenAI now has a model that is simply easier to use and better. Um, when you install it, it doesn't matter what plugins, what stack, whatever, it just starts working. Fable, and now Opus 5— they're a little different. They are a little weirder. They have different ways of working that you have to adapt to. Some will dare to take this step, and some will not. OpenAI is betting that a model that works “out of the box” is the right way to go. And Anthropic is betting that if you create a super- super-brilliant model, everything else will fall into place . Over the past year, it seems to me that Anthropic has had a certain advantage, as if by "mandate of heaven." Um, especially starting with Claude Code about a year or a year and a half ago, they've dominated what is essentially the primary working tool for most people who really love AI. Many people, including me , have started to return to the OpenAI ecosystem. And now Codex is entering its prime . If you look, I think they're adding about a million users a day now. So there's something really interesting going on with OpenAI and their current strategy, and I think the balance of who 's ahead is starting to shift back a little bit towards them. How we like a model, and what we like about it, can sometimes change in a matter of weeks or a month, because you start learning little things, trying it out, and thinking, "Oh, this is a new way to do it." "He's not like my old way." Um, and it's a little annoying that I have to change, but there are hidden possibilities here that I really like. So, um, for us, right now our impression is: some of the cool things are hard to use, and that's not a good thing. However, a change in perception may occur. We've seen this before. Um, I think a very good example of a model that's similar to this—is GPT5, GPT 5.1, GPT 5.2, which were just really difficult to use, and people didn't like them; and about a month later they said: "But it's very powerful for certain things, especially for the CEX line." I think something similar is happening where Fable raised the bar for intelligence so much that it started to spread to all the weaker models, and Anthropic will have to figure out, "How do we best retrain these models to make them more usable, even though they're so smart?" So I would say that chances are you will try this model and you won't really like it. But we'll really know in a few weeks, when everyone will have the opportunity to explore its capabilities to understand what exactly might be hidden there. But so far, to me, this model is a bit like a "poor man's Fable." It has all the personality of Fable, but none of its top-notch features. So for me personally , this is not my favorite model.