If you build with Anthropic models, the interesting part of this release is not the benchmark table — it’s the access pattern. As of September 1, 2026, Anthropic’s Claude Mythos 5.1 is available only to vetted cyberdefenders and life scientists through its verification programs. Everyone else gets Claude Fable 5.1: the same underlying model, wrapped in safeguards tuned to be less annoying without going soft on dual-use work.
Two products from one checkpoint
The distinction matters for anyone planning around model capabilities. According to the announcement, Fable 5.1 can now identify software vulnerabilities in source code — a task the earlier version blocked — and its biology safeguards intervene on benign requests 85% less often than the ones shipped with Fable 5. Hard lines remain: penetration testing, exploit generation, and binary-based vulnerability scanning stay blocked, and dual-use biology and chemistry questions get routed to Opus models as a fallback.
So the decision tree for a builder looks like this: if your org needs unfiltered Mythos-level capability in cybersecurity or bio research, you apply to the Cyber Verification Program or the Life Sciences Verification Program. If you just need strong general capability with fewer false-positive refusals, Fable 5.1 is the practical default. It’s a tiering approach rather than a single public launch.
Pricing, retention, and what the benchmark footnote means
The commercial details are concrete. Mythos 5.1 pricing starts at $10 per million input tokens and $50 per million output tokens, and access requires accepting a 30-day data retention policy by default for safety monitoring. Claude Security now runs on Mythos 5.1, which tells you Anthropic is dogfooding the verified tier in its own product.
The benchmark notes deserve a careful read. Fable 5.1 was evaluated with production safeguards enabled, and on tasks where safeguards intervened, Fable 5.1 scored zero on OSWorld 2.0 while Fable 5 scored zero on AutomationBench — with intervened cybersecurity tasks completed by Claude Opus 4.8 and biology tasks by Claude Opus 5. Anthropic itself flags that this likely understates Fable’s performance. Translation for your evals: a safeguarded model’s measured capability depends heavily on whether your test set trips the tripwires. If you benchmark Fable yourself, expect the same effect.
The broader release history is a cautionary tale
The timeline on the announcement page shows why this pattern exists. Mythos 5 launched on June 9, 2026 to a small group of vetted partners, became unavailable on June 12, and only regained access for a set of US organizations on July 1 after government approval and the lifting of export controls. Version 5.1 followed on September 1, and on October 6 the Cyber Verification Program expanded to cover Opus 5.5, Sonnet 5.5, Mythos 5.1, and future models.
This mirrors what I covered earlier about Anthropic’s three tiers of trusted cyber access: capability is being gated by verified identity and use case, not just by an API key. For teams building security tooling or life-science workflows, the practical step is to start the application process early — these are vetting pipelines, not self-serve toggles.
The takeaway: plan for the Fable tier as your baseline assumption, read the system card before committing to either variant, and treat verified access as a procurement timeline item, not a launch-day dependency. And note the limitation — the figures here come from Anthropic’s own announcement and system card testing, not independent evaluation.