Anthropic launches Claude Fable 5.2 and Mythos 5.2 at $8/$40 per million tokens, claiming a 30-hour Fermat proof
- Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, the same underlying model, at $8 per million input tokens and $40 per million output tokens, with outputs up to 30% faster than the 5.1 versions.
- Prerelease testers used the model for software migrations, designed three working electronic devices from scratch, solved two open problems in black hole physics, and formalized Fermat's Last Theorem in 30 hours.
- Benchmarks put Fable 5.2 above Opus 5.5 on most measures: FrontierSWE v2 at 69.9% against 62.3%, Remote Labor Index at 41.3% against 21.3%, and RoboDojo Sim at 64.9% against 27.6% for Fable 5.1.
- Fable ships with cybersecurity and biology safeguards so all users can access it, while Mythos requires entry into Anthropic's Life Sciences or Cyber Verification Programs.
- An optional fast mode costs $16 and $80 per million tokens and speeds generation up to 2.5 times, and Anthropic still points most coding and knowledge work at Opus 5.5 as the default.
Hacker News opinions
The benchmark table says Fable 5.2 beats Opus 5.5 in almost every category. Was that measured before or after they lobotomized Opus 5.5?
The robotics jump is the wild one, 27.6% to 64.9% on RoboDojo Sim. Everything else moved a few points, that one moved.
Leaks said it would be called Fable 5.5, so either the leaks were bait or they renamed it. And it's already live in my Claude Code, usually there's a rollout delay.
Probably they've got something bigger in the pipeline and that one gets the 5.5 name.
Too expensive for me.
Really? A $200/mo plan gives you an enormous amount of inference on this model.
Cheaper than 5.1 at $8 in and $40 out, and the free resets are unusually generous. Subscription length counts from your first paid bill, even if you canceled and came back.
Claude's writing style is terrible next to GPT or Gemini. It's smart, but the way it talks to you is somehow off.
Opus 5 and Fable 5 were unbearable, cryptic and hard to read. Opus 5.5 is a massive improvement.
Opposite for me. Opus 5.5 is verbose and chatty, I've said 'I'm not reading all that' more than once, and that's with concise mode on.
FrontierCode 2.0 is a bad benchmark for Gemini because it's the only model forced onto Devin CLI while everyone else gets their preferred harness.
Doubling Remote Labor Index to 41.3% while the others sit near 20% is the number I care about. That's the agentic work question.
Those physics papers look like the real thing, not mathematical physics. Odd that the only author listed is Claude and they published on Hexagon instead of arXiv.
Nerfing an existing model right before a launch is the fastest way to wreck your brand. It also happens on literally every release.
Fable 5.2 at high effort matches Opus 5.5 at max effort for roughly half the cost per task, so my bank account gets the final vote.