Context
In this solo How I AI mini episode, host Claire Vo reviews Claude Fable 5.2, Anthropic's new frontier model, after two days of building real things with it. Claire, the founder of ChatPRD, had been openly critical of Claude: she found the original Fable and Opus 5 hard to work with, and had moved most of her coding to Codex. She has since returned to Claude Code and Opus 5.5, and this episode explains what Fable 5.2 adds on top. For PMs who use AI in product work, the value is a concrete, task-by-task picture of where a new model helps, where it does not, and how to route work between two models instead of picking one.
The Big Idea
Fable 5.2 is not a replacement for Opus 5.5. It is a fast, creative, long-running model you reach for when a task needs scope, complexity or creativity, and you pair it with Opus 5.5 where it is weak, especially front end and web design.
Claire's verdict comes from hands-on tests, not benchmarks alone: an animated video from a tweet, flat-pack furniture with no manual, a game, a real pipeline re-architecture, and a 3D typewriter app. The one clear failure was design when the model had no brand or style guide to follow.
Key Insights
Intelligence alone did not win her
Claire's earlier complaint about Claude was never about raw capability. The original Fable was a big leap in intelligence, but she found it "so hard to work with, so annoying to talk to." Opus 5 had the same problem, even when it got the job done.
What changed with Fable 5.2 is the working experience:
- It is fast, both per turn and in how long it works on its own.
- It alerted her when it needed something, "but no more than that."
- She did not hit any of its guardrails in her testing, though she notes that may depend on what she tested.
- She found it funny and enjoyable to work with.
For a PM choosing tools, this is a reminder that how a model collaborates (speed, interruptions, tone) can matter as much as how smart it is.
What Anthropic launched and claims
Claire summarized the launch before her own tests:
- Fable 5.2 ships alongside Claude Mythos 5.2, the same underlying model, available only for cybersecurity and life sciences purposes.
- Anthropic says most daily coding and knowledge tasks are still Opus 5.5 tasks.
- Anthropic's coding claims: strong on stubborn bugs, performance optimization, architecture, long-running tasks and huge migrations.
- On science, the headline is that it solved two open problems in the physics of black holes.
- On benchmarks, Claire said it outpaces the other frontier models at each model's highest effort level, naming Frontier Code, FrontierSuite and Automation Bench.
- Cyber and bio guardrails remain unless you are in the Mythos program, and the model has many anti-distillation safeguards.
Video that stops looking like slides
Claire had been making product marketing videos from customer meeting notes and tweets using HyperFrames with Opus 5.5. She said that combo changed how she does product marketing, but the videos still looked like animated Google Slides or PowerPoint.
- Test: she gave Fable 5.2 one tweet about software factories plus the CXO.dev brand guidelines and said "make this tweet a video."
- Result: a fully animated story with space-invader-style bots, a bug factory and a closing billboard.
- Guidance given: none beyond the brand book. She did no storyboarding.
- Caveat: she did go back and forth with it, and said the result is not a hundred percent perfect, but way better than before.
Strong sense of geometry, motion
Claire's brand had static images of little invader bots. Earlier models only made them wiggle across the screen. Fable 5.2, unprompted, made sprites showing how the bots would move, cheer and jump.
She was impressed because these are abstract objects with no obvious way to move. She liked the result enough to add "animate the invader bots" as a new eval in her How I AI bench, and said she could not get any other model to pull it off.
The flat-pack furniture test
This was the eval Claire trusts most, because she tried it on every model first.
- Setup: dropshipped furniture arrived with an open bag of screws, 25 pieces of wood, Chinese-language labels and no instructions.
- Input: fairly low-resolution photos of the pieces laid out on her floor.
- Before: the previous week, every model she tried failed to produce instructions.
- Fable 5.2: produced a diagram plus step-by-step assembly in the right order, which matters a lot for flat-pack furniture.
- Proof: she and the person who came to assemble it followed the instructions, and the order was correct.
Creative work that is actually funny
Claire extended her CXO bug factory game into a space invader shooter with a short prompt: make a game about fighting bugs using the space invaders from the brand.
- You play an engineer stopping bugs before they reach production, and you pick your challenges.
- More agents stop more bugs, but each missed bug lets more bad stuff through.
- Protected time and tests help you catch bugs.
- The model invented the concepts, tied them to real business ideas, and made a trailer where finance sees the token bill and downgrades the team to a cheap model.
- The final boss is the monorepo, a giant space invader.
She said she played it a few times and giggled. She noted it is not a visually impressive 3D game, but the humor and the business framing were the win.
Real engineering on a messy codebase
Claire pointed Fable 5.2 at the product intelligence pipeline in ChatPRD, which she described as an ingest, organize, summarize "auto wiki" product she has iterated on for about six months, with very messy code.
- Ask: make the pipeline more efficient and easier to maintain, and do its best product thinking.
- Result so far: the work is sensibly organized and matches the roadmap she already had in her head.
- Working style: it was autonomous, stopped at the right moments, and ran until it needed a PR.
- Caveat: she is only about halfway through the migration and will report back on whether it works.
She said this kind of architecture and long-running engineering work is something previous models, such as Opus 5.5 and Astra, had not really been able to get through.
Weak taste without a style guide
The clearest con was front end and web design. With a good brand or style guide, Claire said Fable 5.2 is tasteful and follows it well. Left to its own devices, she said it has "incredibly concerning taste."
- She asked it to design a new browser homepage for a project about slowing down online, and called the result ugly slop.
- A ghostly storybook project came out okay but pretty ugly.
- Her plan is to pair Opus 5.5 for design with Fable 5.2 on the back end.
She also missed a built-in image generation model. Because Fable 5.2 is so creative, she wanted it to create images too, and often had to call OpenAI's image generation model instead. Fable was not very good at that part.
A one-shot 3D typewriter
Claire was inspired by a 3D record player Guillermo Rauch of Vercel built and shared with its source code. She pointed Fable 5.2 at a similar but different problem and got a fully working 3D typewriter, which she says was one-shot, with later tweaks and new features.
- Type by clicking keys or using your keyboard, zoom in, backspace and type over text.
- Music, rain sounds, hourglass writing timers (15 and 30 minutes) and writing prompts.
- Scenes such as a soft evening that follows the time of day, a train, an oil lamp, handwritten pages and a Y2K computer with green or amber screens.
- Print pages to a dot matrix printer, stack them, reopen, save and copy the text.
- Guillermo's record player is built in, with radio stations.
She called it her "AGI moment" and something she could never have imagined building before. It is not perfect: the handwriting and the type-over effect do not always come out right.
Mental Models & Frameworks
Route tasks between two models
Claire does not crown one model. She splits work by strength:
- Fable 5.2: tasks needing scope, complexity or creativity, such as video, animation, 3D, games, geometry puzzles, architecture and long-running engineering.
- Opus 5.5: most daily coding and knowledge tasks (Anthropic's own guidance), plus front end design where Fable's taste is weak.
- A separate image model: for generating images, since Fable 5.2 has none built in.
The principle is that a new model's weaknesses are fine if another model in your stack covers them. In her words, the things Fable 5.2 is not good at, Opus 5.5 is good at.
Build evals from real failures
Claire's most convincing tests came from her own life and work, not public benchmarks:
- A task every model had failed (the furniture), so success meant something.
- A task she had a clear quality bar for (videos that do not look like slides).
- A task she verified in the real world (following the assembly steps).
When a model passes a test like that, she adds it to her How I AI bench, as she did with animating the invader bots. Each new model then gets measured on the same real problems.
Give it brand inputs, not freedom
Across the tests, Fable 5.2 did best when Claire supplied a brand book or style guide and little else. The brand book alone produced a strong video, and the brand's invader sprites drove the game. Without those inputs, the design work fell apart. The pattern: constrain the look, free the creativity.
Trade-offs & Nuance
Fast and autonomous, but unfinished work
Claire praised its autonomy on the ChatPRD migration, but she is only halfway through and has not confirmed the result works. Autonomy that feels good mid-task is not the same as a verified outcome.
Early impressions, not a full benchmark
This is a two-day review. Claire ran a preview of her How I AI bench showing speed (animating the invader bots took 16 minutes, and many tasks ran in under a minute). She plans to run the full bench against other new models in a week or two. Treat the verdicts as strong first impressions.
Not hitting guardrails may be luck
She did not run into any guardrails, but said it might be because of what she was testing. Cyber and bio guardrails still apply outside the Mythos program.
Practical Application
Rerun your hardest failed task
Keep a list of AI tasks that every model has failed for you, like Claire's furniture-from-photos problem. When a new model ships, run those first. A pass on a known failure tells you more than a headline benchmark.
Turn one asset into video
Take one piece of existing content, such as a tweet, a launch note or customer meeting notes, plus your brand guidelines, and ask a creative model to make it a video. Claire did this with no storyboard. Check whether the output looks like animated slides or a real story, and iterate on one sub-loop (like character animation) if needed.
Write down a model routing rule
Decide in writing which tasks go to which model. Claire's version: creative, complex and long-running work to Fable 5.2, daily coding, knowledge work and web design to Opus 5.5, image generation to a separate image model. Share it with your team so people stop debating the "best" model per task.
Always attach a style guide
Before asking Fable 5.2 for anything visual, attach your brand book or style guide. Claire found it tasteful with one and badly off without one.
Judge collaboration, not just output
When trialing a model on a long task, note how often it interrupts you, whether it stops at sensible checkpoints (like before a PR), and how long you wait per turn. Claire's switch back to Claude was driven by these factors as much as by output quality.
Questions to Consider
- Which AI tasks in my product work has every model failed so far, and do I have them written down to test the next model on?
- Where am I using one AI model for everything when a second model would cover its weak spots, such as design or image generation?
- Do my AI-made marketing videos look like animated slides, and have I tried giving a creative model only a brand book and one source asset?
- When an agent runs a long migration autonomously, how will I verify the result actually works before I trust its architecture?
Bottom Line
Claire Vo's verdict is that Fable 5.2 is a fast, creative, enjoyable model that cracked tasks no other model could, from flat-pack assembly to a one-shot 3D typewriter. It still has poor design taste without a style guide and no image generation, so the smart move is to route creative and complex work to Fable 5.2 and keep Opus 5.5 for daily tasks and design.
Case Studies Mentioned
Tweet to animated video
Claire gave Fable 5.2 a tweet about software factories and the CXO.dev brand guidelines, with no storyboard. It produced a full animated story about a team building a software factory that fails because the repo is not ready, checks are slow and nobody is trained. The lesson: a strong creative model with only brand inputs can move AI video past the animated-slide look.
ChatPRD pipeline re-architecture
Claire asked Fable 5.2 to make ChatPRD's six-month-old product intelligence pipeline more efficient and easier to maintain. The plan it followed was sensibly organized and matched her own unwritten roadmap, and it worked autonomously until it needed a PR. The migration is about halfway done and not yet verified.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Claude Fable 5.2 | Anthropic's new frontier model | The subject of the review, praised for speed, creativity, geometry and long-running coding, criticized for web design taste. |
| Claude Mythos 5.2 | The same model as Fable 5.2, for cybersecurity and life sciences use | Launched alongside Fable 5.2; the Mythos program lifts cyber and bio guardrails. |
| Claude Opus 5.5 | Anthropic's model for most daily coding and knowledge tasks | Claire's main coding model, and her pick to pair with Fable 5.2 for design work. |
| Claude Code | Anthropic's coding agent | Where Claire used Fable 5.2, and where she now does most of her coding. |
| Codex | OpenAI's coding agent | Claire's favorite coding tool before she returned to Claude Code, especially for browser use and computer use. |
| HyperFrames | A tool for making videos with AI | Claire used it with Opus 5.5 to turn notes and tweets into product marketing videos. |
| ChatPRD | Claire's AI product for PMs | Its product intelligence pipeline was her real engineering test for Fable 5.2. |
| How I AI bench | Claire's own eval set for AI models | She previewed Fable 5.2's speed on it and will run the full bench against other models. |
People to Follow
Claire Vo
Host of How I AI and founder of ChatPRD. She reviews new AI models by building real projects with them and maintains her own How I AI eval bench, which makes her reviews useful for PMs picking tools.
Guillermo Rauch
The Vercel leader whose 3D record player, shared with its source code, inspired Claire's typewriter project. Claire built his record player into her typewriter app.
Experiments to Try
Experiment: Brand book only video
- Hypothesis: a creative model given only your brand guidelines and one source asset will produce a video that feels less like animated slides than your current AI video workflow.
- Change: feed Fable 5.2 one tweet or launch note plus your brand book, with no storyboard or creative direction.
- Test: compare the output side by side with a video from your current tool on the same source.
- Measure: how many revision rounds each needed, and whether teammates can tell which one looks like slides.
Experiment: Design with and without guide
- Hypothesis: Fable 5.2's design quality depends heavily on whether it gets a style guide, as Claire observed.
- Change: ask for the same landing page twice, once with your brand or style guide attached and once without.
- Test: run the same prompt through Opus 5.5 as a comparison.
- Measure: which version your designer would ship with the fewest changes.
Notable Quotes
"Intelligence does not make you intelligible." (Claire Vo)
"It is not a replacement for Opus 5.5." (Claire Vo)
"I did not give it any creative guidance other than our brand book." (Claire Vo)
"The things that it is not good at our buddy Opus 5.5 is good at." (Claire Vo)
