Anthropic Ships Claude Fable 5: Mythos-Class Power With Guardrails, Crushing Every Coding Benchmark
Anthropic released Claude Fable 5, a publicly available version of its restricted Mythos-class model, scoring 80.3% on SWE-Bench Pro — 11 points above Opus 4.8 and 21 above GPT 5.5. Risky cyber and bio queries automatically fall back to Opus 4.8.
Anthropic released Claude Fable 5 on June 9 — the most capable model it has ever made generally available, and effectively a public version of the restricted Mythos-class systems it had previously reserved for government partners and vetted cyberdefenders.
The benchmark gap is not subtle. Fable 5 scores 80.3% on SWE-Bench Pro, against 69.2% for Anthropic’s own Opus 4.8, 58.6% for GPT 5.5, and 54.2% for Gemini 3.1 Pro. On FrontierCode, the hardest public coding eval, it hits 29.3% — more than double Opus 4.8’s 13.4% and five times GPT 5.5’s 5.7%. Stripe reported that Fable 5 completed a codebase-wide migration in a single day that an engineering team had scoped at over two months.
The interesting part is the safety architecture. Fable 5 ships with AI classifiers that monitor three risk areas: cybersecurity, biology and chemistry, and model distillation. When a query trips a classifier, the response is silently generated by Opus 4.8 instead. Anthropic says the fallback triggers in under 5% of sessions, and admits the filters are “too aggressive” in places, occasionally catching harmless requests. It is a novel trade: frontier capability for everyone, with the sharpest edges routed to a weaker model.
The unfiltered version exists too. Claude Mythos 5 — the same underlying model with safeguards lifted in specific areas — goes to a small group of cyberdefenders and infrastructure providers through Project Glasswing, Anthropic’s partnership with the US government. On ExploitBench, Mythos 5 scores 78%, up from 69% for Mythos Preview and 40% for Opus 4.8. In molecular biology, its hypotheses were preferred over Opus-class output 80% of the time in blinded comparisons, and Anthropic claims a 10x speedup in parts of drug-design workflows. A Trusted Access Program for biology researchers is planned.
Pricing is steep but lower than expected: $10 per million input tokens and $50 per million output — nearly double Opus 4.8, but less than half what Mythos Preview cost. Claude subscribers get Fable 5 included at no extra charge through June 22; from June 23 it consumes usage credits at a 2x rate. The model is live on the Claude API and Amazon Bedrock today.
The release lands two days after Anthropic’s confidential S-1 filing at a $965 billion valuation. Shipping your best-ever model the same week you court public-market investors is not a coincidence — it is the pitch. Anthropic spent months warning that frontier AI is getting dangerous, then released its most powerful model anyway. The classifier-fallback system is the company’s answer to that tension, and the rest of the industry will be watching whether it holds.