Anthropic: Claude Fable 5 and Mythos 5 — the most powerful models ever released

On June 9, 2026, Anthropic released Claude Fable 5 (claude-fable-5) and Claude Mythos 5 (claude-mythos-5), models that according to both internal and external benchmarks surpass every previously released Claude model. Both belong to the Mythos class — a new capability tier above Opus — but are intended for different user groups.

What are Fable 5 and Mythos 5?

Claude Fable 5 is a Mythos-class model adapted for general use, with a full set of safety protections. Claude Mythos 5 shares the same underlying architecture but with relaxed guardrails for authorized users — currently exclusively Project Glasswing partners, a collaboration between Anthropic and the US government for applications in cyber defense and critical infrastructure.

Both models support:

• 1 million tokens context window by default
• 128,000 tokens maximum output
• Always-on adaptive thinking — cannot be disabled

Benchmarks and capabilities

Fable 5 achieves the highest scores on FrontierCode for mid-strength programming, the Hebbia Finance Benchmark for financial analysis, and CursorBench for coding assistance. In Pokémon FireRed it completed a full playthrough using a minimal, vision-only role, with no special adapter layers.

In drug biological design applications, internal experiments show approximately 10× acceleration in protein molecule design: of 14 tested protein targets, 9 produced strong candidates for further research. Mythos 5 additionally shows that scientists preferred its hypotheses ~80% of the time in blind comparisons against alternative models, and one hypothesis was confirmed in an independent bioRxiv study. On genomics tasks, the model autonomously ran a single-cell analysis across 138 species and built a custom ML model 100× smaller than the reference Science journal model — which nonetheless outperformed it.

The company’s demo integration with Stripe describes how Fable 5 “compressed months of engineering work into days” when migrating a codebase of 50 million lines of Ruby code.

Tokenizer and pricing

Fable 5 and Mythos 5 use the tokenizer introduced with Claude Opus 4.7. The same text produces approximately 30% more tokens compared with models predating Opus 4.7 — development teams should re-measure prompts using the token-counting API with the parameter model: "claude-fable-5".

Fable 5 API pricing:

• Input tokens: $10 per million
• Output tokens: $50 per million

That is less than half the price of the previous Claude Mythos Preview model.

Safety architecture and classifiers

How do refusals work?

Fable 5 runs safety classifiers on every request and during response generation. When a classifier refuses a request, the Messages API returns stop_reason: "refusal". There is no charge if the refusal occurs before any output has been generated.

The new stop_details.category field now includes three refusal categories:

• "cyber" — offensive cyber tasks and exploit development
• "bio" — bioweapons and dual-use research
• "reasoning_extraction" — blocks model distillation (prevents replication of Fable 5 capabilities)

The optional fallbacks parameter (in beta on the Claude API) allows automatic retry of a refused request on another model, billed at that fallback model’s rates.

On average, classifiers activate in fewer than 5% of sessions.

What differs for Mythos 5 users?

Mythos 5 is available exclusively to Project Glasswing partners — initially for cyber defenders and critical infrastructure providers working with the US government. It is an upgrade from the previous Claude Mythos Preview model.

Mythos 5 requires 30-day data retention on the Claude API and is not available under zero data retention. All human-operator access to human data is logged; data is not used for model training.

Mythos 5 retains all “hard-limit” class safety protections, but with relaxed boundaries for legitimate cybersecurity and biology use cases.

Red-teaming and jailbreak resistance

Anthropic conducted external bug bounty testing covering more than 1,000 hours of testing — without uncovering any universal jailbreaks.

One external partner described Fable 5’s cyber protections as “the most robust of any model tested” and confirmed that the model “complied with harmful single-step requests zero times” across 30 jailbreak techniques.

UK AISI (AI Security Institute) made progress toward a universal jailbreak in a “brief initial testing window” on agentic tasks — the only known exception in red-team reports.

Availability

Plan	Access	Conditions
Pro, Max, Team, Enterprise	June 9–22, 2026	Included at no extra charge
Pro, Max, Team, Enterprise	From June 23, 2026	Requires usage credits
API users	From June 9, 2026	Standard per-token billing
Mythos 5	Immediately	Project Glasswing partners only

Claude Fable 5 marks the first Mythos-class release for the general public and establishes a new reference point for the capabilities of large language models available to commercial users.