Anthropic Launches Claude Fable 5: Frontier AI Capabilities Now Public

Key Development

AnthropicReleased Claude Fable 5 on June 9, 2026, making Mythos-class capabilities available to the general public for the first time. The release pairs a general-access model (Fable 5) with a restricted variant (Mythos 5) reserved for vetted cybersecurity and critical infrastructure operators.

Fable 5 represents a deliberate split: the same underlying model, but with safety classifiers that automatically route high-risk queries to Claude Opus 4.8 instead. Mythos 5, offered through Project Glasswing in collaboration with the US government, removes those safeguards for trusted defenders.

Capability Benchmarks

Fable 5 leads across nearly all measured domains. On SWE-Bench Pro—solving real software engineering tasks from public repositories—it scores 80.3%, compared to Claude Opus 4.8 at 69.2%, GPT 5.5 at 58.6%, and Gemini 3.1 Pro at 54.2%. The model excels at extended, multi-day autonomous tasks, with early testing showing Stripe used Fable 5 to compress months of engineering into days.

Performance gaps widen on complex, long-horizon work. For shorter, straightforward prompts, the advantage over Opus 4.8 narrows—but for ambitious projects spanning days and requiring sustained reasoning, Fable 5’s lead is material.

Safety Design and Trade-offs

AnthropicTook an unusual approach: ship the same model twice, split by access controls rather than capability caps. Fable 5 blocks requests in cybersecurity, biology, chemistry, and model distillation—falling back to Opus 4.8 silently. In internal testing, these safeguards trigger in fewer than 5% of sessions, meaning 95% of Fable 5 interactions run entirely on the model’s own responses.

All Fable 5 and Mythos 5 traffic requires 30-day retention for safety monitoring. Anthropic says the data is not used for training, only to identify novel attacks and reduce false positives.

During red-teaming, Mythos Preview (the predecessor) identified and exploited zero-day vulnerabilities in every major operating system and web browser when directed to do so. That capability exists in Fable 5 too—but only Mythos 5 can exercise it openly.

Pricing and Availability

Both models cost $10 per million input tokens and $50 per million output tokens—less than half the price of Mythos Preview. Fable 5 is available immediately via the Claude API, AWS Bedrock, Vertex AI, and Microsoft Foundry. Subscription access rolls out in stages: through June 22, Pro, Max, Team, and seat-based Enterprise plans include Fable 5 at no extra cost. From June 23, usage credits apply.

Industry Context

The launch comes as Anthropic prepares for an expected IPO and signals a shift in AI release strategy: frontier models can ship publicly if engineered carefully. It also underscores the coding-first competition among labs. Anthropic has reclaimed the #1 spot on major coding benchmarks—a market segment worth billions in enterprise spend.

Open Questions

How effectively will Fable 5’s classifiers hold against novel jailbreaks in production? Will competitors match the two-model split strategy, or maintain single-track releases? And how will the 30-day retention requirement affect enterprise contracts in jurisdictions with strict data localization rules?

For builders, the practical question is simpler: does Fable 5’s edge on your specific workload justify the cost premium over Opus 4.8? Early testing suggests the answer is yes for complex coding and long-context reasoning—but less clear for simpler tasks.


Source: Anthropic