Published: June 26, 2026

Anthropic Launches Its Most Powerful Models and Turns Them Off Three Days Later

On June 9, Anthropic released Claude Fable 5 and Claude Mythos 5. Three days later, both were disconnected due to a U.S. export control directive.

Detalle del proceso

On June 9, Anthropic launched Claude Fable 5 — the most capable model made available to the public to date — alongside Claude Mythos 5. Three days later, both were taken offline. If you read the announcement and considered trying them, here’s what you need to know beforehand.

What was launched

Fable 5 is a Mythos-class model designed for long and complex tasks: code migrations, extensive analysis, workflows where the model operates for hours without losing focus. It features a context window of 1 million tokens and output of up to 128,000 per request, at $10 per million input tokens and $50 per million output tokens — double that of Opus 4.8.

Mythos 5 is the same model internally, but with some security filters lifted. It is not available to the general public: it is offered only to a select group of organizations through Project Glasswing, in collaboration with the U.S. government. If you're not in that program, the Mythos-class model for everyone else remains Fable 5.

The twist: both are suspended

On June 12, the U.S. government issued an export control directive requiring the shutdown of Fable 5 and Mythos 5 for any foreign individuals, inside or outside the country. Since Anthropic could not filter by nationality quickly enough, it deactivated both models for all clients worldwide. The rest of the models — including Opus 4.8 — continue to operate normally. Anthropic claims this is a misunderstanding and says it is working to restore access. As of today, with no set date.

What changes if you integrate them

When they return, behaviors will differ from Opus 4.8. Adaptive thinking is always active: it cannot be turned off, and the model decides how much to reason based on the task; if you need to control that depth, use the effort parameter. It does not return raw chain-of-thought — only a summary, and only if you request it. Additionally, in Fable 5, cybersecurity, biology, or chemistry queries are not answered by the model itself: they are routed to Opus 4.8. This occurs in less than 5% of sessions, but it’s worth considering. And when there is an explicit rejection, the API returns stop_reason: "refusal" with an HTTP 200, not an error: your code must treat it as a valid response from the start. One more point: both are Covered Models, with 30-day data retention and no zero-retention option. Something to keep in mind before integrating them into workflows with sensitive data.

Our perspective

The capability of Fable 5 is real — maintaining judgment across hundreds of chained decisions unlocks things that previously weren’t possible: long UX audits, coherent content systems at scale, automation with production weight. But the news this week isn’t what the model can do, but that you can’t use it.

For now, the actionable advice is this: if you had plans for Fable 5, develop them on Opus 4.8, which remains available, and prepare the migration for when, and if, access is restored.

Read more news

More recent articles to keep exploring the blog.