Anthropic has unleashed its latest flagship model, Claude Fable 5 — and it’s the first time the company has opened a Mythos-level AI to the public. Pricing starts at $10 per million input tokens and $50 per million output tokens, more than half off the beta price. A sibling model, Claude Mythos 5, also launched with fewer safety restrictions, aimed at security-focused organizations.
On FrontierCode’s Diamond-level benchmark, Claude Fable 5 scored 29.3% — blowing past GPT-5.5’s 5.7% and Claude Opus 4.8’s 13.4%. In another test, Stripe used the model to migrate a 50-million-line Ruby codebase in a single day; the same task used to take a team two months.
Visually, the model can play through Pokémon FireRed just by analyzing screenshots — and then reconstruct the web page source code from those same screens. In biology, Fable 5 predicted Dyno Therapeutics’ adenovirus shell assembly through pure reasoning, outperforming specialized protein models. It also independently designed and trained a single-cell genomics classifier that beat the one published in Science, despite using only 1% of the parameters.
To keep these capabilities from being abused, Anthropic deployed a separate safety classifier that automatically falls back to Claude Opus 4.8 for high-risk queries — affecting fewer than 5% of sessions. Enterprise traffic data will be retained for 30 days, and any human review gets logged. For Glasswing partners, Claude Mythos 5 already has cybersecurity restrictions removed, and biomedical limits will gradually open to specific institutions via a trusted access mechanism.
The new models are fully available through the API and enterprise tier. Subscription users can try them free until June 22nd; after that, they’ll need to pay with consumption credits until Anthropic has enough compute to fold them back into the regular subscription.