Anthropic Releases Claude Fable 5, Its Most Capable Public Model Yet

iEXExchanger
Anthropic Releases Claude Fable 5, Its Most Capable Public Model Yet

Anthropic released Claude Fable 5 — its most capable widely available model, with Mythos-level power but built-in safeguards. Alongside it comes Claude Mythos 5, unrestricted but only for vetted customers.

Anthropic has officially released Claude Fable 5 — and by the company's own account, it's the most capable model it has ever made widely available. The "Mythos-class" model that leaks were circling just yesterday is now open to everyone, in a safe version.

Two Claudes instead of one

Anthropic unveiled two models of the same capability tier at once. Claude Fable 5 is the public one: it runs on Anthropic's API as well as Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. Claude Mythos 5 has the same power but without the safety classifiers; it isn't open to all, and is offered only in limited release through the Project Glasswing program to vetted customers, mainly for cybersecurity and life-sciences work. Put simply, Fable 5 is Mythos made safe for everyone.

What it can do

The main bet is on long, autonomous tasks. The model can run complex coding and analysis work for hours — something earlier versions couldn't sustain for so long. It reads visuals well — diagrams, charts, and tables in files and PDFs, useful for research, finance, and legal work. And it checks itself: it adjusts skills as it goes and builds its own test routines. The context window reaches 1 million tokens, with up to 128,000 tokens in a single response. Per Anthropic, the model is state of the art on nearly every benchmark tested.

When it refuses, Opus 4.8 answers

Fable 5 has built-in safety classifiers: for some requests — say, on cyber weapons, biology, chemistry, or health — it may refuse. It's handled cleanly: instead of an error you get a normal response flagged as a refusal, and the request can be routed automatically to another model, Opus 4.8. You aren't charged for a refusal that generates no output. Two more details: "adaptive thinking" is always on, and the raw chain of thought isn't returned — only a short summary.

The price is $10 per million input tokens and $50 per million output. Both models are tagged as Covered Models with 30-day data retention. But the bigger point is this: what was a closed checkpoint under a codename a day ago is a public product today. The race for the strongest model has entered a phase where new releases land back to back.

Questions and answers

Frequently asked questions about this article

How does Claude Fable 5 differ from Mythos 5?

They share the same capability level, but Fable 5 has built-in safety classifiers and is available to everyone. Mythos 5 has no such limits but isn't open — only in limited release through Project Glasswing for vetted customers.

What does 'Mythos-class' mean?

It's the top capability tier in Anthropic's lineup. Claude Fable 5 reaches it but with built-in safeguards, while the fully unrestricted version of that power is available only to a narrow set of customers as Claude Mythos 5.

What does Fable 5 cost and where is it available?

It's $10 per million input tokens and $50 per million output. The context window reaches 1 million tokens. It's available via Anthropic's API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry from June 9, 2026.

What happens when the model refuses a request?

Instead of an error you get a normal response flagged as a refusal, and the request can be routed automatically to another model, Opus 4.8. You aren't charged for a refusal that produces no output.