Claude Fable 5 is the most capable AI model Anthropic has ever released widely. It opens the Mythos class, a tier above the Opus line, and targets coding or research projects that run for days on end. The same brain ships two ways: Fable 5, safeguarded and broadly available, and Mythos 5, with filters lifted, reserved for vetted cyberdefenders. Access goes through a paid Claude plan or the API, billed $10 per million input tokens.
Claude Fable 5 is built for long, ambiguous, asynchronous work: it plans in stages, delegates to sub-agents and checks its own code. Its context window reaches 1 million tokens, with 128,000 output tokens per request, and it reads text, images and files.
The showcase story comes from Stripe, which migrated a 50-million-line codebase in a single day. In practice, you hand it the ticket you would have scoped at a week, switch on auto mode in Claude Code and review the result over breakfast. Thinking stays on at all times (an effort dial sets how hard it reasons, and how fast the bill grows).
Fable 5 and Mythos 5 share the same underlying model; only the safeguards differ. Fable 5 carries safety classifiers: a sensitive request touching cybersecurity, biology or chemistry gets declined or handed to Opus 4.8 as a fallback. Mythos 5, stripped of those filters, stays confined to Project Glasswing, run with the US government.
There was a twist along the way: export controls forced Anthropic to cut access a few days after launch, before a worldwide return three weeks later. A rare episode in the short history of large language models.
| Criterion | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Access | Broad: API, cloud, Claude plans | Restricted to Project Glasswing |
| Safeguards | Active classifiers, Opus 4.8 fallback | Lifted in specific areas |
| Audience | Developers, companies, subscribers | Vetted cyberdefense organizations |
| Underlying model | Identical | Identical |
Fable 5 comes with Claude Pro, Max, Team and Enterprise plans, on Claude.ai as well as in Claude Code and Claude Cowork. Developers call it under the claude-fable-5 model ID from several platforms.
On price, the API charges $10 per million input tokens and $50 for output, less than half the old Mythos Preview rate, with a 90% discount on prompt caching. US-only inference runs at 1.1x those figures. These numbers shift with each announcement, so Anthropic's pricing page has the final say.
No, Claude Fable 5 requires a paid Claude plan (Pro, Max, Team or Enterprise) for the chat apps, and API calls are billed by token volume at a higher rate than the Opus line. Free Claude accounts do not include it; Sonnet remains the entry point there.
Access was cut when the US government applied export controls to Fable 5 and Mythos 5 shortly after launch. With no reliable way to verify nationality in real time, Anthropic paused the model for everyone, then restored worldwide access once the controls were lifted a few weeks later.
It depends on task length. Claude Opus 5 lands close to Fable 5 on many jobs at roughly half the cost, but Fable 5's lead widens as work gets longer and more complex. Multi-day agent runs and big migrations justify it; routine chats sit fine on Opus.
Anthropic requires a 30-day retention window for all traffic on Mythos-class models, Fable 5 included, so misuse patterns can be analyzed after the fact. That sits on top of the classifiers screening requests during the exchange itself. Companies with strict confidentiality rules should factor this window into their compliance review.
Verdict: Giant code migrations, dense research files, agents that must last the distance: that is Fable 5 country. Developers and research teams who kept bumping into Opus-class ceilings will find the headroom they were waiting for.
GPT-6 Astra 
Claude Fable 5.1 
Grok 4.6 
