Fable 5 is online again
On July 1, 2026 Claude Fable 5 came back. Nineteen days after being switched off by order of the US government.
Let's start from the beginning, for anyone who missed the earlier chapters. On June 9 Anthropic released Fable 5, the most powerful model ever built. Three days later, on the 12th, the US Commerce Department imposed export controls and Anthropic had to shut it down for everyone, worldwide. We covered it here: the day the US government switched off Fable 5.
Now the end of the story: on June 30 the restrictions were lifted, and from July 1 Fable 5 is accessible again on Claude.ai, Claude Code, Claude Cowork and via the API. Re-enablement on AWS, Google Cloud and Microsoft Foundry is underway.
What changed to bring it back
Two things. One technical, one political.
On the technical side, Anthropic built a new safety classifier aimed precisely at the bypass technique reported by Amazon — the one that triggered the block. According to the company, the filter stops that specific exploit in over 99% of cases.
On the political side, the government gave the green light. Commerce Secretary Howard Lutnick stated that his team worked with Anthropic to analyze and approve Fable 5, to align positions and strengthen America's leadership in AI.
There's one detail that matters more than all the rest. Testing revealed that the disputed capability could be reproduced with other models too — GPT-5.5, Opus 4.8, even Haiku 4.5. In other words: the alleged risk was not unique to Fable 5. It's the technical confirmation of what Anthropic had said from the start, when it spoke of a misunderstanding.
On what terms you can use it now
The return isn't a fully open tap, at least for now.
On subscription plans — Pro, Max, Team and select Enterprise — through July 7 Fable 5 is included up to 50% of weekly usage limits. After that date, access moves to additional usage credits. Via the API the model (model ID: claude-fable-5) is available on a pay-as-you-go basis, at launch pricing: $10 per million input tokens, $50 output.
Mythos 5, the version without some protections reserved for the government Glasswing program, has only partially returned: restricted access for a group of US organizations, with the expansion still in progress. For almost every company, though, the model that matters is Fable 5 — and it's back.
Want an AI adoption that holds up even when models change?
30 minutes to discuss your specific case.
What you actually do with it, now that you can
It's worth remembering why Fable 5 made noise, beyond the political saga.
On coding it's at the top. Stripe migrated a 50-million-line Ruby codebase in a day — a job that would have taken over two months with traditional methods. On financial reasoning Hebbia gave it the top score on its benchmark, with Bloomberg calling it senior-researcher level. On vision it set the state of the art: it extracts numbers from scientific figures, reconstructs code from a screenshot, reliably reads tables inside a scanned PDF.
And it holds up on long context: it keeps the thread across millions of tokens in extended sessions, with file-based memory that multiplies performance on genuinely long tasks. We put the technical details here, in the launch article: Fable 5 and Mythos 5, the most powerful model ever built.
Translated for a business: code migrations, document analysis over hundreds of pages, compliance audits across entire archives, data extraction from scanned documents. Things that were a workaround with earlier models become reliable with Fable 5.
When Fable 5 is worth it, and when it isn't
Fable 5 isn't the model to use for everything. It's the frontier model: expensive, powerful, right for the hardest tasks.
For the bulk of daily work — agents at volume, automations, repetitive document processing — the workhorse remains Claude Sonnet 5, which delivers near-top-tier quality at a fraction of the price. You keep Fable 5 for where you need the maximum: the deepest reasoning, the longest task, the analysis where an error costs. The right logic is to route — which task to which model — as we explain in Opus, Sonnet and Haiku compared.
With Fable 5 back on the menu, the choice widens upward. But the principle doesn't change: you pay for the flagship model only where the result justifies it.
The lesson that remains
There's one thing this story leaves behind, beyond the happy ending.
A production model, inside the workflows of hundreds of millions of people, was switched off and back on by a government decision within nineteen days. Not because of a bug. Because of a political choice.
It's the reminder we keep giving our clients: adopting AI doesn't mean marrying a model. It means building a system that holds up even when a piece moves — deprecated, updated, or switched off by government order. Those who had abstracted the model away from their processes, in June, absorbed the interruption without stopping. Those who had tied everything to a single name did not.
Fable 5 is back. The right question isn't "do I use it or not," but "do my processes hold if tomorrow it gets pulled again." If you want to understand how to build an AI adoption that doesn't depend on a single model, let's talk.