Anthropic released Claude Fable 5 today, the first public-facing version of its Mythos model. The model delivers strong performance on software engineering, knowledge work, and vision tasks, but comes with hard safety limits baked into the inference pipeline. In high-risk domains—cybersecurity, biology, chemistry, and distillation—the model falls back to Opus 4.8, effectively gating responses at the classifier level.

Mythos debuted as a preview in April, restricted to a handful of partners due to security concerns. Last week, Anthropic expanded access to hundreds of critical-infrastructure organizations across 15 countries. Now, any developer can hit the Claude API or consumption-based Enterprise plans. Subscription users get Fable 5 included at no extra cost through June 22; after that, it shifts to usage credits until Anthropic reinstates it as a standard feature.

Anthropic also released Mythos 5 for pre-approved organizations. Both models are priced at $10 per million input tokens and $50 per million output tokens—double Opus 4.8. That sticker shock may be intentional: enterprises are already burning through annual AI budgets on advanced reasoning models that split single requests into multi-step chains.

Safety architecture is front and center. Anthropic ran an external bug bounty that logged over 1,000 hours of testing without producing a universal jailbreak. Red teams also came up empty. Still, the company is mandating 30-day traffic retention for all users, even those with prior zero-retention agreements. Data won’t be used for training, only for detecting novel attacks and reducing false positives. Expect this to become an industry pattern: access to powerful models increasingly tied to mandatory retention policies.

Deferrals to Opus 4.8 are rare—early data shows at least 95% of Fable sessions run entirely on the model. Third-party benchmarks support the performance claims: Hex reported Fable hit 90% on its core analytics benchmark, and Base44 noted strong one-shot app generation and tool-calling. Genspark found it beat all competitors on UI design and game coding.

Anthropic expects unpredictable demand. Rakuten, an early adopter, sees the extra thinking as paying for itself. For engineers evaluating this, the trade-off is clear: higher per-token cost and mandatory data retention in exchange for a model that self-validates and rarely gatekeeps—unless you’re working on something that could actually cause harm.

Source: TechCrunch