Fable 5 is back: what it is and what you can do with it

july 1, 2026 · selma kocabıyık

Launched three weeks ago, shut down by the US government three days later, back online today. In this piece I tell the story of Fable 5 — Anthropic's most powerful public model — what it is, and what you can practically do with it.

I'm @selmaaii — I publish content about Claude Code and AI workflows. Follow along for more.

The quick news first: what happened?

Anthropic released Claude Fable 5, its first public Mythos-class model, on June 9, 2026. Three days later, on June 12, the US government issued an export-control directive on national-security grounds and the model was shut down worldwide. The reason: Amazon researchers reported a method that bypassed Fable 5's safety measures (getting the model to find software vulnerabilities — in one case producing code demonstrating an exploit).

Then the balance shifted. On June 30 the controls were lifted, and on July 1 (today) Fable 5 came back globally.

The return isn't unconditional: the model came back more cautious, with a new safety classifier. It blocks the technique Amazon reported at over 99% and reroutes blocked requests to Opus 4.8 — and this new filter can flag harmless requests more often, even in routine coding/debugging work. On cloud providers (AWS, Google Cloud, Microsoft Foundry) access is rolling out gradually.

What is Fable 5?

Fable 5 is the first publicly available example of what Anthropic calls "Mythos-class" models — a tier above Opus. The biggest difference is autonomy: the model no longer just answers a single question; it plans and executes long, multi-step tasks on its own. It takes a goal, decides which tool to use, uses it, makes mistakes, and recovers from them by itself. The old mindset was "answer me"; Mythos is "finish this job".

There's also a sibling model: Mythos 5. The same underlying model, but with some of the safety measures removed. It's available only to a limited set of organizations in Project Glasswing and carries the strongest cyber and science capabilities. Think of Fable 5 as "Mythos power + public safety measures". One practical consequence: questions in risky areas like cyber, biology and chemistry get downgraded to Opus 4.8 on Fable 5 (in under 5% of average sessions).

Technical summary: a 1-million-token context window, up to 128K output tokens per request. The API model string is claude-fable-5.

What can you do with Fable 5?

This is the real question. Fable 5's strength isn't "short one-off questions"; it's long-running, context-heavy work that needs to be finished end to end. Concretely:

Large code migrations and hard software work

In early testing, Stripe completed a migration on a 50-million-line Ruby codebase — one that would take a team more than two months — in a single day. Fable 5 scores 80.3% on SWE-Bench Pro, ahead of Opus 4.8 (69.2%) and GPT-5.5 (58.6%). In other words: for the hardest work in complex, real codebases.

Tools, games and apps from scratch

It coded its own browser-based CAD editor (VibeCAD), then designed a 3D-printable model inside that tool. It played Factorio autonomously and built a fully automated factory from scratch. So beyond "write me code", it constructs a working system end to end.

Visual understanding

It finished Pokémon FireRed with no map or game-state data — screenshots only; it can rebuild a web app's source code from screenshots alone. It's strong anywhere you need to get work done from visual inputs: screenshots, diagrams, tables, charts.

Long-context data and document work

With 1M-token context plus persistent memory, it doesn't forget what it learned during a long task. In the Slay the Spire test, file-based memory improved its performance 3× more than it did for Opus 4.8. In practice: reading long reports and drawing conclusions, turning data into action items, multi-document research flows.

Long research flows

On the science side, the most striking results were shown on the sibling model, Mythos 5: speeding up drug design ~10× (strong candidates for 9 of 14 protein targets), and a genomics model trained on single-cell data from 138 animal species that beat a model published in Science while being 100× smaller. Note: these ran on Mythos 5 with safety measures removed; on Fable 5, biology queries may get downgraded to Opus 4.8, so don't expect the same directly. But it shows where the level of autonomy has reached.

How do you get access, and what does it cost?

As of July 1, three routes are open:

• claude.ai / Claude Code / Claude Cowork: pick Fable 5 from the model menu.
• API: model string claude-fable-5.
• Cloud: access on AWS, Google Cloud and Microsoft Foundry is rolling out gradually; it may not be ready on day one — keep a fallback model configured.

Pricing: on Pro, Max, Team and select Enterprise plans, Fable 5 is included up to 50% of your weekly usage limit until July 7; after that it's metered via usage credits. On the API it's $10 per million input tokens and $50 per million output tokens. Fable 5 is the most expensive model in the lineup — not "the model for everything" but "the specialist for hard work". For daily chat, writing and mid-level coding, Opus 4.8 is still more economical; save Fable 5 for genuinely hard, multi-step jobs.

What should you watch out for?

• These are controlled demos. Pokémon, Factorio and the Stripe test are impressive but staged scenarios; they don't guarantee the same result on your messy data. Validate with a small test.
• The new classifier is more cautious. Even in routine coding/debugging you can get blocked and rerouted to Opus 4.8; this is the most visible change the return brought.
• The science results are on Mythos 5. The drug and genomics demos were done on Mythos 5 with safety measures removed, not on Fable 5.
• There's 30-day data retention. Because Fable 5 and Mythos 5 count as "Covered Models", prompts and outputs are retained for up to 30 days for safety classifiers, and zero-data-retention (ZDR) is not offered. Factor this in for sensitive data.
• Pricing changes on July 7. The included window is short; if you want to try it, start before that date.

TL;DR

• Fable 5 was shut down by government order on June 12 and returned on July 1 — with new safety measures.
• What it is: the first public Mythos-class model above Opus — autonomous, finishing long tasks on its own.
• What it does: large code migrations, tools/games/apps from scratch, visual understanding, long-context data and research work.
• Evidence: Stripe 50M lines/1 day, SWE-Bench Pro 80.3%, 3× memory gains; the science demos are on the Mythos 5 side.
• Cost: the most expensive model ($10/$50), included up to a 50% limit on plans until July 7; save it for the hard jobs.

If you've tried something with Fable 5, share your results on X — let's see together where it genuinely makes a difference.