Anthropic has released Claude Fable 5 and Claude Mythos 5, the strongest Claude models to date. Fable 5 is open to all users; Mythos 5 is limited to trusted users. Within 24 hours of launch it topped the Agent Arena leaderboard, setting the record for the largest margin in AI history.

What the Two Versions Are

• Claude Fable 5: the version with a safety net, open to all users. When a risk classifier trips (cybersecurity, biochemistry, and the like), it automatically downgrades to Opus 4.8 to answer
• Claude Mythos 5: the full-power unrestricted version, used only by a small set of trusted users

The two are versions of the same underlying model. Base technical metrics are completely identical; the difference lies only in the safety guardrails.

Core Capabilities

Software Engineering

On the SWE-bench Pro evaluation, Fable 5 scored 80.3% versus GPT-5.5's 58.6%. In Stripe's 50-million-line Ruby codebase, Fable 5 completed a full-codebase migration that would normally take two months in a single day.

On Cognition's Frontier Code evaluation, Fable 5 took the highest score among frontier models at medium reasoning effort.

Frontend Development

Highlights from developer hands-on testing:

• Liquid glass CSS: recreating Apple's WWDC26 liquid-glass effect in pure CSS, no WebGL involved
• Windows on the web: login, notifications, settings, a full Solitaire suite, with a built-in Copilot
• An 8000-line Pokemon clone: built in one pass, recreating all 151 original Pokemon
• A national park with 266,000 trees: real elevation data and satellite imagery, navigable in the browser

Vision

Fable 5 achieves "native vision, no scaffolding." From game screenshots alone, with no external aids of any kind, it played its way through Pokemon FireRed entirely on its own.

Long Context and Memory

It stays focused across million-token, long-horizon tasks and uses file-based memory to improve its output. With persistent file memory attached, its Slay the Spire performance improved 3x.

Research Capability

In biomedicine, Mythos 5 can independently execute a complete biologist workflow. Of the 14 protein-target complexes it designed, 9 have entered real drug-development pipelines in the lab.

Price and Limits

That's double Opus 4.8's pricing, and actual token consumption is larger too. Wharton professor Ethan Mollick noted after testing that tokens burn very fast.

Limited-time free window: through June 22, Pro, Max, Team, and Enterprise users can use Fable 5 for free. From June 23 it requires additional usage credits. API and pay-as-you-go enterprise customers are unaffected by this cadence.

Safety Mechanism

Behind Fable 5 sits a set of independent classifiers that check whether a request involves cyberattacks, biochemical risk, or model distillation. When one trips, the request is automatically downgraded to Opus 4.8 to answer, rather than flatly refused.

Over 95% of sessions never trigger a downgrade. But the classifiers lean conservative: legitimate tasks such as biologists studying viruses or security engineers running authorized attack-defense exercises can also get caught. Anthropic says it will bring the false-positive rate down going forward.

How to Get It

• Web: claude.ai, choose the Fable 5 model
• API: called through the Anthropic API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry
• Claude Code: use it through Claude Code, which supports Dynamic Workflows for multi-agent parallelism

Usage Advice

1. Now through June 22 is the free window — try it while you can
2. Start with small tasks; token consumption is far higher than expected
3. It suits complex, long-horizon tasks rather than simple single-turn Q&A
4. Frontend development, code migration, and long-document analysis are the standout scenarios today