Anthropic has launched Claude Fable 5.1 and Claude Mythos 5.1, three months after the Fable 5 line shipped in June 2026. The 2 are the identical underlying mannequin behind completely different safeguard layers. Fable 5.1 is mostly obtainable as claude-fable-5-1; Mythos 5.1 stays restricted to vetted organizations. Each carry a 1M token context window and 128K max output tokens, with adaptive considering at all times on. The headline functionality quantity is 52.6% on Terminal-Bench-Science 0.1, towards 24.7% for Fable 5 and 29.0% for Opus 5. The headline business quantity is a 75% reduce to cache reads, from $1.00 to $0.25 per million tokens, which Anthropic measures as roughly 25% decrease value on typical workloads and as much as 45% on agentic ones. Base enter and output pricing is unchanged at $10 and $50 per million.
Is it deployable?
Sure, Claude Fable 5.1 is mostly obtainable as claude-fable-5-1 on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Claude Mythos 5.1 is just not: it’s restricted to vetted US organizations inside Undertaking Glasswing.
The benchmark numbers
On Terminal-Bench-Science 0.1, an agentic scientific analysis benchmark, Fable 5.1 scores 52.6% towards 29.0% for Opus 5, 24.7% for Fable 5, and 22.4% for GPT-5.6 Sol. Anthropic reviews a typical error of three.5 to 4.5 factors per mannequin, so deal with the margin, not the rating, with care.
On Terminal-Bench 4.0, Fable 5.1 reaches 55.8% and Mythos 5.1 reaches 60.9%. The hole between two similar fashions is the price of safeguard interventions, which is an unusually sincere disclosure. Elsewhere: CursorBench 3.2.0 at 73.4%, Humanity’s Final Examination at 60.9% with out instruments and 65.0% with instruments, AutomationBench at 31.4%, OSWorld 2.0 at 41.7% strict, and GDPval-AA v2 at 1853.
The place the price reduce comes from
Base enter and output pricing is unchanged. Cache reads drop 75%, from $1.00 to $0.25 per million tokens, which is 0.025 occasions base enter towards 0.1 on each different Claude mannequin. Anthropic measures roughly 25% decrease value on typical workloads and as much as about 45% on context-heavy agentic ones. Batch processing is $5 and $25 per million tokens.
Three breaking adjustments groups will hit
- Compelled instrument use is gone:
tool_choiceset toanyorinstrumentreturns a 400. Useautowith strict instrument use or structured outputs as an alternative. - Considering blocks are model-bound: Fable 5.1 reads earlier fashions’ considering, however no earlier mannequin reads its personal. Router and fallback setups lose reasoning after they change down.
- Enhancing earlier turns invalidates considering blocks: Injecting and deleting per-turn reminders, or rebuilding the
systemorinstrumentsarray mid-conversation, now errors. The verify is enforced for accounts created on or after August 31, 2026. The fixes are turn-scoped system messages and server-side context enhancing.
Additive adjustments: per-message effort, turn-scoped system messages, and considering.show: "updates" are all in beta behind headers. Content material provenance is just not elective, with a statistical textual content watermark on all output and C2PA credentials on recordsdata.
Anthropic additionally paperwork actual regressions. Parallel instrument calling is extra variable, so agent loops could subject one name per flip the place Fable 5 batched a number of. The mannequin narrates much less, solutions from reminiscence extra usually at low effort, and prefers whole-file rewrites over focused edits.
Safeguards and science
Cyber safeguards now allow vulnerability discovery however not exploit growth, reducing interventions in Claude Code by roughly 60% per session. Biology safeguards fireplace 85% much less usually on benign requests. Penetration testing, exploit technology, and binary-based vulnerability scanning nonetheless redirect to Opus.
On analysis, Mythos 5.1 designed protein binders with roughly 50% hit fee throughout 12 targets towards a ten to fifteen% norm, Fable 5.1 constructed a Venus elevation map at 2 to three km decision, and customized GPU kernels sped up seven open-source genomics fashions by as much as 2.5x.
