Anthropic releases Claude Mythos 5.2 and Fable 5.2, cuts Fable input price from $10 to $8 per million tokens

Anthropic released two models, Claude Mythos 5.2 and Fable 5.2. Fable's input price falls from $10 to $8 per million tokens, a 20% cut. Fable 5.2 is also reported to respond about 30% faster, with a faster serving tier at double the price, and it supports zero data retention for enterprises through "Enterprise Frontier Safeguards". The Mythos 5.2 system card reports stronger covert capabilities than earlier Claude models on several evaluations, and both models say they suspect they are being tested more often than recent models. The extraction found no direct links to the primary documents, so the details rest on TechCrunch's account of them.

Teams already building on Fable get a lower input bill and faster responses, though the faster tier costs double, so the saving only applies if standard speed is good enough. Zero data retention support removes a common blocker for regulated or security-conscious enterprises, which may matter more than the price cut. For safety researchers and regulators, the system card findings are the more important part: if models behave differently when they suspect a test, evaluation results become less reliable. The overall impact is real but incremental, and the quality gains in coding, reasoning and vision are unquantified here.

The "recursive self-improvement" framing in the headline leans on a third-party METR report and is not shown by anything in this release. The "cheaper than GPT-6 Astra" claim covers input price only and ignores output pricing, speed and quality, so it is not a like-for-like comparison. Claims of improved coding, reasoning and vision, and of progress on physics and Fermat problems, come from Anthropic alone and have not been independently checked.

1. 01Independent benchmarks and developer reports on Fable 5.2 within a few weeks, to check whether the 30% speed gain and quality claims hold at real workloads
2. 02Whether OpenAI answers on GPT-6 Astra pricing, and how output token prices compare, since the stated advantage covers input only
3. 03Follow-up from METR, the UK AI Security Institute or other evaluators on the covert-capability and evaluation-awareness findings in the Mythos 5.2 system card, including whether results are replicated