Claude Fable 5 and Claude Mythos 5
Anthropic News · 2026-06-09
Anthropic launches Claude Fable 5, a state-of-the-art general-release model with new safety classifiers that fall back to Opus 4.8 on high-risk queries, alongside Claude Mythos 5, the same underlying model with safeguards lifted for vetted cybersecurity and biomedical research partners.
Appears in
Extraction
Topics: large-language-modelsai-safetyanthropic-claudeagentic-aifrontier-ai-models
Claims
- Claude Fable 5 sets state-of-the-art benchmarks across software engineering, vision, knowledge work, and scientific research, with the largest performance gains on longer and more complex tasks.
- Fable 5 includes safety classifiers that fall back to Claude Opus 4.8 for queries related to cybersecurity, biology and chemistry, and model distillation, triggering in under 5% of sessions.
- Claude Mythos 5 is the same underlying model as Fable 5 but with cyber safeguards lifted for vetted partners in Project Glasswing and select biomedical researchers.
- Mythos 5 autonomously conducted novel genomics research over more than a week, training a model that outperformed a recently published Science journal paper while being 100 times smaller.
- Anthropic requires 30-day data retention for all Mythos-class model traffic to monitor for complex attacks, jailbreaks, and false positives.
- Both models are priced at $10 per million input tokens and $50 per million output tokens, less than half the price of Claude Mythos Preview.
Key quotes
Fable 5's capabilities exceed those of any model we've ever made generally available.
In a 50-million-line Ruby codebase, the model performed a codebase-wide migration in a day that would otherwise have taken a whole team over two months by hand.
Mythos 5 conducted novel genomics research in over a week of largely autonomous work... Mythos 5's trained model outperformed a recent model published in the journal Science—despite being 100 times smaller.
We've deliberately tuned the safeguards to be cautious, and they are still stricter than would be ideal—for example, sometimes benign requests will trigger our classifiers.