Claude Fable 5: Anthropic's most powerful model

What happened

On 9 June 2026 Anthropic opened general access to Claude Fable 5 — its most powerful publicly available model to date. Fable 5 belongs to the Mythos class: it is the same engine as the restricted Claude Mythos 5 model, but with added safety classifiers that make a broad release possible. Mythos 5, without those guardrails, is available only to selected partners under the Project Glasswing program.

Key specs (per Anthropic and the platform documentation):

• Price: $10 per million input tokens and $50 per million output tokens.
• Context: a 1M-token window by default, up to 128k output tokens per request.
• API identifier: claude-fable-5.
• Coding: the highest score among frontier models on the Cognition FrontierCode and FrontierBench evaluations.
• Agentic tasks: it works autonomously for longer than any previous Claude model. Stripe reported that the model "compressed months of engineering work into days" and migrated a 50-million-line Ruby codebase in a single day.
• Vision: claimed as the new state of the art on image tasks — up to reconstructing a web app's source code from screenshots alone.

An important detail for developers: Fable 5 has built-in safety classifiers (cybersecurity, bio/chem, distillation attempts) that can reject a request. On a refusal the Messages API returns stop_reason: "refusal" as a successful HTTP 200 rather than an error, and suggests falling back to another model (fallback to Claude Opus 4.8). Per Anthropic, the protection fires in fewer than 5% of sessions on average.

Then the adventures began. On 12 June the US government applied export controls to Fable 5 and Mythos 5 — after Amazon researchers found a way around Fable 5's safety mechanisms. Since verifying users' citizenship in real time was impossible, Anthropic temporarily cut off access for everyone. The export restrictions were lifted on 30 June, and on 1 July Fable 5 returned globally — in Claude Platform, Claude.ai, Claude Code and Claude Cowork. For the return, the company added an improved classifier that blocks the technique Amazon described in more than 99% of cases.

Why it matters

First, this is Anthropic's sharpest jump in raw capability in some time — Fable 5 is positioned above the Opus/Sonnet/Haiku line and, on internal benchmarks, sets a new bar in coding, agentic scenarios and image work.

Second, the release context is unusual. Per TechCrunch, only days before the launch Anthropic publicly warned that frontier systems are advancing so fast they could approach recursive self-improvement, and called on labs to coordinate on safety measures. Shipping the most powerful model right after such a warning raised obvious questions — and the company answered them with layers of defense: classifier stress tests, a bug bounty and, by its own account, no universal jailbreaks across 1,000+ hours of testing.

Third, this is the first high-profile precedent of US export controls effectively killing access to a frontier model for every user in the world for three weeks, including the company's own foreign-national employees. For the industry it's a signal: the availability of top models now depends not only on the labs but on regulators too.

Worth noting separately is the data policy: Fable 5 and Mythos 5 carry 30-day traffic retention (Covered Models), including for customers who previously had zero-retention agreements — Anthropic explains this as protection against sophisticated attacks.

What it means for vibe coding

For anyone coding with AI, Fable 5 is a noticeable upgrade, but with caveats:

• Long agentic tasks got more realistic. A 1M-token window and the ability to run autonomously longer than earlier models hits vibe coding's sore spot directly: big refactors, migrations and multi-step tasks without constant babysitting.
• Be ready for refusals. If you're building an integration on claude-fable-5, handle stop_reason: "refusal" and plan a fallback (server-side via the fallbacks parameter, client-side via SDK middleware, or manual). Anthropic grants a fallback credit so you don't pay twice for cache when switching models, and doesn't bill a request rejected before generation.
• The price is for serious work. $10/$50 per million tokens is markedly more than the Sonnet class, so it's cheaper to keep smaller models for routine work and switch Fable 5 on for the hard parts. On paid plans after its return, Fable 5 was included within 50% of the weekly limit through 7 July, and via usage credits after that.
• Factor in availability. The export-control episode showed that a frontier model can become unavailable overnight. For production it's sensible to keep a backup path to Opus 4.8 or another model.

The practical takeaway: Fable 5 is a tool for heavy, long, agentic tasks where the quality gain justifies the price; leave everything else to cheaper models and always plan a fallback.

Sources

• Claude Fable 5 and Claude Mythos 5 — Anthropic
• Redeploying Claude Fable 5 — Anthropic
• Introducing Claude Fable 5 and Claude Mythos 5 — Claude Platform Docs
• Anthropic released Claude Fable 5, days after warning AI is getting too dangerous — TechCrunch
• Anthropic releases Mythos-like AI model to the public, Claude Fable 5 — CNBC
• White House lifts export control on Anthropic that froze its most advanced models — CNN Business