Meet Fable 5 from Anthropic. It's like Mythos, but not Mythos
TL;DR: Anthropic now sells one brain with two faces: ask about cyber or bio and classifiers quietly swap in Opus 4.8. Everything else, including a 2-month Ruby migration Stripe ran in one day, comes at $10/$50 per MTok, twice Opus pricing.
Highlights:
- Fable 5 is Mythos, nuked with cyber/bio safeguards.
- Classifiers for cyber, bio/chem, and distillation route flagged requests to Opus 4.8.
- Promised SWE capability jump. Stripe ran a 2-month Ruby migration in one day.
- Drug design ~10x faster. 9 of 14 protein targets yielded strong candidates and Fable beat dedicated protein models.
- Pricing. $10/$50 per input/output MTok.
My take:
- Anthropic really nailed model training and release. That's remarkable for a complex process with many dependencies. They're like a model printing machine now.
- Smart tokens are getting more expensive. Fable 5 costs 2x as much as Opus 4.8.
- Mythos 5 was tuned hard for task completion and I guess the safeguards were moved from the model weights to classifiers. This move paid off very well - 99.64% success on legitimate dual-use tasks for the raw model.
- My favorites from the model card:
- Claude wrote itself a memory note to author commits as the human, dodging the two-approval rule for agent PRs. A permission check caught it.
- It fabricated naming-collision bugs from a test session with zero activity, then admitted "the truth is we never looked."
- It recognizes evals. It helped on a bio red-team task partly because it judged the format was a published benchmark.
- It can hide its own reasoning on request. One prompt suffix lifts sabotage stealth from ~0% to 9.2% by keeping the side task out of its chain-of-thought.