Anthropic launched Claude Fable 5.1, its most advanced model for coding and knowledge work, GA on all platforms (claude-fable-5-1 on the API, plus AWS, GCP, Azure) and made it the default Fable model in Claude Code. Pricing keeps $10/M input and $50/M output but cuts cache reads from $1.00/M to $0.25/M, making typical workloads ~25% cheaper than Fable 5 and highly cache-heavy agentic workloads up to ~45% cheaper. Benchmarks: Terminal-Bench 4.0 55.8% (Fable 5 42.0, GPT-5.6 Sol 37.3), Terminal-Bench-Science 0.1 52.6% vs GPT-5.6 Sol 22.4, GPQA 88.5 (new SOTA), HLE 60.9 without tools. The same model ships as Mythos 5.1 for trusted-access programs (bio via LSVP with the US government, cyber via CYP), with stronger cyber capabilities but below Anthropic's next bio risk tier. Anthropic also announced Enterprise Frontier Safeguards (ZDR-equivalent privacy with safeguards, phases starting fall 2026) and, for models released after Aug 2 2026 (Fable 5.1 included), output watermarking per the EU AI Act with a detection API in private preview.
For anyone running agentic workloads on Claude, the cache-read cut is the headline: agents that re-read long frozen context every turn are exactly the profile that saves up to ~45%, which changes model-routing math for coding agents specifically. The benchmark spread on Terminal-Bench 4.0 (55.8 vs GPT-5.6 Sol's 37.3) plus effort levels now exposed on a 1-10 scale makes effort-based cost control practical. Watch the two governance shifts: Enterprise Frontier Safeguards converges on OpenAI's Private Safety Processing (retention-free inference with safety tooling), and the EU watermark + detection API turn provenance marking into an integration requirement for European deployments.
| Availability | claude-fable-5-1 GA on all platforms: Claude Platform, Claude Code, Claude Enterprise, AWS Bedrock, Google Cloud, Microsoft Azure; default Fable model in Claude Code as of 2.1.257 |
|---|---|
| Pricing | $10/M input, $50/M output (unchanged); cache reads $0.25/M (down 75% from $1.00/M); typical workloads ~25% cheaper than Fable 5, highly agentic workloads up to ~45%; 1M context window; 1-10 effort scale (public), defaults High in Claude Code / Medium in Claude Cowork and Claude.ai |
| Coding Agent Benchmarks | Terminal-Bench 4.0: Fable 5.1 55.8%, Mythos 5.1 60.9% vs Fable 5 42.0 / Opus 5 52.3 / GPT-5.6 Sol 37.3; Terminal-Bench-Science 0.1: 52.6% vs Fable 5 24.7 / Opus 5 29.0 / GPT-5.6 Sol 22.4; CursorBench 3.2.0: 73.4% vs GPT-5.6 Sol 65.8 / 71.2; OSWorld 2.0 strict: 41.7% (prior open-weights SOTA 24.3); AutomationBench: 31.4% |
| Knowledge Benchmarks | GPQA 88.5 (new SOTA); Humanity's Last Exam 60.9 without tools / 65.0 with tools; AIME 2026 98.9; ARC-AGI-2 36.1 |
| Mythos 5 1 | same base model with stronger safeguards posture, distributed only via trusted access: Life Sciences Vulnerability Program (LSVP, with the US government) and Cyber Vulnerability Program (CYP); strongest cyber capabilities of any Anthropic model, placed lower-risk in the Frontier Compliance Framework, below the next bio risk tier in the RSP |
| Efs | Enterprise Frontier Safeguards: ZDR-equivalent privacy with safeguards; customer-controlled cloud storage; phased availability starting fall 2026 (eligible customers get ZDR until then); supported on Claude Code, Claude Enterprise, Claude Platform, Bedrock, Google Agent Platform, Foundry |
| Cyber Safeguards | ~60% fewer false-positive interventions vs Fable 5; Fable 5.1 allowed for vulnerability discovery (defensive), exploit development redirected to Opus models; Mythos-class customers get the enhanced set |
| Watermark | EU AI Act Article 50(2) applies to models placed on the market after 2026-08-02 — Fable 5.1 outputs watermarked; watermark detection API in private preview |
| Anti Distillation | new API accounts can no longer edit prior context while preserving Claude's thinking transcript (closes a documented distillation technique) |
| Science | protein binder design ~50% experimental hit rate across 12 targets (typical 10-15%), 10x binding affinity vs the best competition entry on 3 targets; CC-licensed Venus elevation map covering ~1/3 of the planet at 2-3 km resolution (prior global ~10-20 km); 7 open-source biology DL models accelerated up to 2.5x with custom GPU kernels (30-60% GPU cost cut), kernels to be open-sourced |
| Cross Ref | companion alignment/security report published 2026-08-31 (ev-20260831-01); EU watermark mechanism background in ev-20260814-07 |