Claude最强Fable 5.1发布！8项屠榜，最高降价45%，反蒸馏机制上线
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - Anthropic released Claude Fable 5.1 and restricted-access Mythos 5.1, claiming benchmark-leading performance, stronger multi-step reasoning, and workload savings of up to 45%. The release also introduces signed reasoning blocks that prevent developers from altering conversation history while retaining prior chains of thought.
- Fable 5.1 reportedly ranks first on eight public benchmarks, with notable gains in scientific and coding tasks; Mythos 5.1 targets approved cybersecurity and life-sciences organizations.
- Cache-read pricing falls 75% to $0.25 per million tokens, while input and output prices remain $10 and $50; Anthropic estimates roughly 25% lower typical costs and up to 45% savings for highly agentic workloads.
- Demonstrations include experimentally tested protein binders, higher-resolution Venus terrain mapping, and GPU kernels that accelerated seven computational-biology models by up to 2.5× with identical outputs.
- New cryptographic signatures bind reasoning blocks to the original system prompt, tools, and message history; modified prefixes can trigger an HTTP 400 error or cause invalid reasoning blocks to be dropped.
Sources (1)
Claude最强Fable 5.1发布！8项屠榜，最高降价45%，反蒸馏机制上线
TL;DR - Anthropic released Claude Fable 5.1 and restricted-access Mythos 5.1, claiming benchmark-leading performance, stronger multi-step reasoning, and workload savings of up to 45%. The release also introduces signed reasoning blocks that prevent developers from altering conversation history while retaining prior chains of thought.
- Fable 5.1 reportedly ranks first on eight public benchmarks, with notable gains in scientific and coding tasks; Mythos 5.1 targets approved cybersecurity and life-sciences organizations.
- Cache-read pricing falls 75% to $0.25 per million tokens, while input and output prices remain $10 and $50; Anthropic estimates roughly 25% lower typical costs and up to 45% savings for highly agentic workloads.
- Demonstrations include experimentally tested protein binders, higher-resolution Venus terrain mapping, and GPU kernels that accelerated seven computational-biology models by up to 2.5× with identical outputs.
- New cryptographic signatures bind reasoning blocks to the original system prompt, tools, and message history; modified prefixes can trigger an HTTP 400 error or cause invalid reasoning blocks to be dropped.