Claude Fable 5.2 review: what the launch-day testers actually found

What Anthropic actually shipped

Two names, one model. Claude Fable 5.2 is the generally available version, open to all paying users on Claude's paid plans, in Claude Code, via the API (model ID claude-fable-5-2) and on AWS, Google Cloud and Microsoft Azure. Claude Mythos 5.2 is, in Anthropic's own words, "the same underlying model" with some cybersecurity and biology safeguards removed, restricted to vetted organizations. Unless you run a security or biomedical lab, Fable is the one on your menu.

The standard API price is $8 per million input tokens and $40 per million output — 20% below Fable 5.1's $10/$50, according to Decrypt's launch coverage — with an optional fast mode at $16/$80 that generates up to 2.5× faster, and batch pricing at $4/$20 per KULIMO's comparison. Output runs up to 30% faster than 5.1. Even so, Fable costs twice as much per token as Claude Opus 5.5 — and Anthropic says so itself: "For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver." A company steering most users away from its newest flagship is unusual. The testers, it turns out, mostly agree with it.

Anthropic's headline claims, all vendor-reported and not yet independently verified: a record 49.1% on its FrontierCode 2.0 coding evaluation at the Xhigh effort level (Fable 5.1 scored 32.4%); first place on the independent Artificial Analysis Intelligence Index at 65 points, ahead of Gemini 4 Carbon (60) and Opus 5.5 (59); roughly 1,000 Fable agents formalizing a proof of Fermat's Last Theorem in 30 hours (formalizing an existing proof, not discovering one — Opus 5.5 managed 26% in 18 hours with the same setup, per Tiago Sampaio's launch-day analysis); three working electronic devices designed from scratch; a prototype operating system built by 1,000+ subagents over eight days; and a 64.9% average success rate on the RoboDojo two-arm robot simulation, more than double Fable 5.1's 27.6%.

One more launch detail matters for everything that follows: Fable has an effort dial (Low, Medium, High, Xhigh, Max). Anthropic's own figures show the dial matters more than the badge — at High effort, Fable roughly matches Opus 5.5's best FrontierCode score at about half the cost per task, and at Low it matches Opus's top score on Humanity's Last Exam for under a tenth of the cost. The record lives at Xhigh. The value lives lower down.

The consensus in one word: spiky

The most useful launch-day verdict comes from Dan Shipper at Every, who tested Fable 5.2 across coding, research, design and writing and landed on a single organizing idea: the model is "spiky." Extraordinary on a narrow set of ambitious, long-running tasks; the wrong tool for almost everything routine. His scorecard, as reported by BigGo's write-up of the Every review:

• Coding: S-plus. "Changes what is possible in software engineering."
• Knowledge work: S for the right person, B otherwise. Transformative if you restructure how you work around it; slow, expensive and token-hungry for email and documents.
• Design: A. "A new ceiling for models," with a warning to tune the effort dial.
• Writing: B to B-plus. Readable, but it buries the lead — behind Gemini 4 Argon and Astra.
• Overall: S — "paradigm-shifting but spiky."

Shipper's measure is worth quoting because every other tester echoes it: "The best measure of a new model is whether it causes you to try things that you never would have tried before. And by that measure, Claude Fable 5.2 is a true paradigm shift of a model." Every's head of evals, Mike Taylor, put the routing rule more memorably: Fable is "a rare senior colleague that you want to use deliberately," while Opus feels like "a $30 an hour Upwork contractor" you can use for everything.

What the testers built

The launch-day evidence is not benchmarks; it is finished software, much of it built from prompts short enough to fit in a sentence.

A video editor, twice over. At Every, Kieran Klaassen built EveryCut — a full video editor combining features the team previously took from Premiere, Descript and StreamYard — end to end by prompting, over a couple of weeks, for thousands of dollars in tokens against what Shipper estimates would normally be millions in investment. Independently, Forward Future's Matthew Berman reports his producer Alex McCraw built a Premiere-style editor the same way, and has been cutting real videos in it since.

A Photoshop of one. McCraw also built Safelight, a Photoshop-style editor configured around his own shortcuts and habits rather than Adobe's. He opened a 4K document, made a YouTube thumbnail with layers, brushes and text — and has since used Safelight for several of the channel's real thumbnails. Berman's verdict in his Forward Future field test: we're entering an era of "software for one" — and McCraw has cancelled the Adobe subscription he held for a decade. For anyone who edits images professionally, this is the launch story that should hold your attention. The moat being tested is no longer the software; it is the workflow knowledge of the person describing what they want. Our own corner of that world — the AI Image Editor and the thirteen models behind it — sells capability by the image; Fable sells the workshop itself.

Games and spreadsheets at datacenter prices. Berman's headline builds: a playable Gears of War–style level from a three-line prompt — about 8.5 billion tokens, roughly $15,000 at API rates by his estimate, with Decrypt tallying $14,710 over 43 hours 48 minutes — and a browser Excel clone with formulas and pivot tables that took ~50 hours and an estimated $11,700. A Fall Guys–style game ran about $13,000. His summary is unqualified: "Fable 5.2 is the best AI model I've ever used, by far." He is equally clear that the results weren't flawless — awkward animations and bugs included.

450 worlds from a paragraph. Wharton professor Ethan Mollick, who discloses pre-release access with unmetered tokens and no lab funding, describes Portals: one paragraph of prompt, an endless journey through code-drawn 3D worlds — 204 worlds after 20 hours, past 450 since, across an estimated 25,000 agent runs, as catalogued in Sulat's roundup of early testers. In his strangest test, Fable read the control software of a small light-up USB keyboard — for which, as far as he could tell, no public instructions exist — and wrote six games for it.

The creative tests. On Lenny's Podcast, Claire Vo spent two days building with it: an animated marketing video from a single tweet and a brand book ("doesn't look like slides"), assembly instructions for flat-pack furniture derived from blurry floor photos with no manual — a geometry test she says every other model failed — and her CXO brand sprites extended into a full bug-defense game with a trailer. She also re-architected a real product-intelligence pipeline with it. Her reservation is specific and relevant here: she still wouldn't let it design a website on its own.

The bill is the real review

Read the launch coverage in sequence and a pattern emerges: the capability question is settled, and the argument has moved to the meter. Fable's token price fell 20% this generation, but its appetite grew faster than the discount. The flagship demos consumed billions of tokens because the model works for days, testing and fixing its own output while the meter runs. Shipper's research documentary — one explainer video, cut from a recorded talk — consumed about 1.4 billion tokens: roughly $2,000 at Fable 5.1 rates, or $862 at Opus 5.5 rates. His conclusion is economic, not technical: at these prices, this class of workflow belongs to companies and the "token rich," though he expects the same workflows to become very cheap within six months to a year.

The other end of the scale matters too. Berman's motion-graphics explainer cost about $25. A volume slider built in ten minutes reportedly cost $8.43. Fable is not uniformly expensive; it is expensive exactly in proportion to how long you let it run unsupervised. Which makes the practical skill of this model generation the same one this studio is built around: route by job, and never pay flagship prices for draft work. It is why our model guide keeps coming back to the same workflow — draft at 1 credit, finish at flagship prices — and why Anthropic's own advice (start on Opus, climb the effort dial one step at a time) reads like the same sentence in a different currency.

Where it falls short

Writing. The clearest consensus weakness. Shipper scores prose B to B-plus: readable, even emotionally aware, but prone to burying the lead and to what Every's Katie Popiel called the manner of "a smug know-it-all." Gemini 4 Argon is, in Shipper's words, "by far better" at writing, with Astra a close second. If your work is words, the flagship is not the frontier.

Web design. Lenny's two-day test flags the same soft spot from the visual side: strong at systems and geometry, not yet trustworthy as an unsupervised web designer. Design at the A tier (Shipper) means interfaces and detail fidelity — Socrates correctly given a black beard for his age in the Protagoras — not taste you can ship blind.

Routine work. Every tester who touched email, documents or quick questions reached the same conclusion: it is, in Shipper's phrase, "way too powerful" — slower and hungrier than the job justifies. And the showcase builds came with disclosed caveats: Berman notes bugs and unfinished features in his, several testers worked with free or unmetered tokens, and most launch-week results have not been independently re-checked. Treat the miracles as promising, not proven.

What this means if you make images

Full disclosure of scope: Fable 5.2 is a language and agent model, not an image model. It will not appear in our model lineup, and nothing in this review is a claim about image quality. What it changes for image people is the layer above the models:

• Your tools are becoming negotiable. Safelight and EveryCut are the same event from two directions: a professional describing their workflow precisely enough got a personal version of a flagship application. The durable skill is the description — the brief, the references, the acceptance test. That is prompt craft at application scale, and it transfers directly from the discipline in our prompt guide.
• Agents will drive image pipelines. A model that works unsupervised for days, checks its own output and fixes its mistakes is, in production terms, an art director who can also operate the studio: brief it, and it can generate, select, edit and re-generate. The models in this studio already expose the pieces; Fable-class agents are the missing foreman.
• The economics rhyme. Fable's lesson — capability is abundant, metered autonomy is the cost — is the image market in miniature. Per-image prices have collapsed to a cent; what you pay for now is iteration, selection and finishing. Expect the same spikiness: spectacular on the big brief, wasteful on the thumbnail.

Our verdict

Claude Fable 5.2 is the strongest generally available model of this launch cycle, and the first one whose reviews read less like benchmark reports and more like org charts: senior colleague, not appliance. Use it the way its own maker and its toughest testers converge on — hand it the long, open-ended, previously-impossible job, set the effort dial deliberately, and watch the meter. Keep Opus 5.5 (or your current daily driver) for everything interactive, routine or written. And if you build creative tools or pipelines, pay closest attention to Safelight: the most consequential thing built with Fable this week was not a game that cost $15,000, but an image editor one person now uses instead of Photoshop.

Sources

• Anthropic — Claude Fable 5.2 and Claude Mythos 5.2 (launch post, Nov 9, 2026; vendor-reported benchmarks and claims)
• Every / Dan Shipper — launch-day review, via BigGo: "S-plus on coding but not at the frontier for writing"
• Matthew Berman / Forward Future — I Put Fable 5.2 to Work: 5 Real-World Field Tests
• Decrypt — Impressive, But Be Prepared to Sell a Kidney to Afford Them (cost tallies, pricing, subscriber access)
• Lenny's Podcast / Claire Vo — Claude Fable 5.2 review: fast, creative, and finally not annoying
• Tiago Sampaio — Claude Fable 5.2: When the Strongest Model Is Worth It (effort-dial analysis)
• Sulat — Claude Fable 5.2 in Claude and Devin: what early testers built (Ethan Mollick's projects)
• KULIMO — Claude Fable 5.2 Compared: Benchmarks & Pricing (Artificial Analysis index, batch pricing)