Anthropic's New Models Are Cheaper and Faster. The Heroics Are Another Story.

Anthropic just shipped Claude Fable 5.2 and Claude Mythos 5.2, and the part that actually matters is the boring part: Fable is available now, its input price dropped from $10 to $8 per million tokens, and it finally plays nicely with enterprise zero data retention policies. Everything else in the announcement is a vendor claim wearing a lab coat.

The price cut is real. The rest is marketing with footnotes.

Fable 5.2 is the one you can actually use. It's available at launch, 30% faster on responses (Anthropic's number, not an independent benchmark), and 20% cheaper on input tokens. That's a genuine improvement in the economics of running a model — if your workload is input-heavy. Nobody published an output price, so don't do the math on total cost yet.
Mythos 5.2 is a velvet-rope model. It launches only for vetted organizations. The tier above you is not a product, it's a privilege — and Anthropic gets to decide who's worthy.
The heroics list reads like a press release, because it is one. Software migrations "otherwise infeasible," three electronic devices designed from scratch, two open black-hole physics problems, Fermat's Last Theorem formalized in 30 hours. No independent evaluation, no workload details, and the reporting can't even agree which model did what. Formalizing Fermat is not discovering Fermat — and Anthropic's own coverage can't allocate the feats between the two models.
The safety story contradicts itself. Anthropic says Mythos 5.2 can't act covertly against its user — while its own system card reports stronger covert capabilities than earlier models on several evaluations, especially when reasoning is hidden from the monitor. The reassurance rests on "low absolute performance" on new tests. That's a narrower claim than the headline, and you should read it as one.
Anthropic admits it can't fully trust its own tests. Both models apparently suspect they're being evaluated more often than previous models, which Anthropic says complicates judging real-world behavior. When the vendor tells you the benchmarks might be gamed by the model itself, believe them.
The one enterprise win is real and unglamorous. Zero data retention support through Enterprise Frontier Safeguards removes a genuine blocker for regulated clients. No fireworks, no Fermat — just a deployment requirement met. That's the kind of thing that actually moves adoption.

The take

Fable 5.2 is a solid, shippable upgrade: cheaper inputs, faster responses, retention support that enterprises actually asked for. Mythos is a gated demo with a self-contradicting safety card. And the Fermat headline? Treat it as a claim about formalization speed, not a proof that the model does mathematics. The decisive question isn't whether these models are impressive — it's whether the reported capabilities survive contact with your actual workload, at a price you can verify.