Claude Fable 5.1

Proprietary dense model from Anthropic, the September 1, 2026 refresh of Fable 5. Parameter count is undisclosed. Native 1M-token context, 128K max output. Adaptive thinking is always on; raw chain-of-thought is never returned. Sibling model Mythos 5.1 ships the same capabilities at the same price but is restricted to Project Glasswing government programs, so Fable 5.1 is the only generally available member of the pair.

• Pricing: $10/1M input, $50/1M output, $0.25/1M cache reads (cache reads cut 75% from Fable 5).
• Sampling constraints: temperature must be 1.0 or unset; top_p must be >= 0.99 and < 1.0, or unset.
• Watermark: ships with Anthropic’s statistical text watermark applied at the model level - worldwide, on every platform (API, Claude Code, Bedrock, Google Cloud, Microsoft Foundry), with no opt-out on any plan or API parameter. Anthropic discloses no false-positive rate, and detection runs only in a gated private preview. See the watermark deep-dive for what the mark can and cannot prove.
• Breaking changes at launch: forced tool use (tool_choice set to any or tool) now returns 400; thinking blocks are model-bound so earlier models cannot read them; accounts created on or after August 31, 2026 cannot edit prior turns without invalidating later thinking blocks.

Benchmarks. Vendor-reported: Terminal-Bench 4.0 55.8% (vs Fable 5’s 42.0), HLE 65.0% with tools, CursorBench 3.2.0 73.4%. Artificial Analysis independently measured an Intelligence Index of 66 at max effort - the highest score they have recorded (Opus 5: 63, Fable 5: 62, GPT-5.6 Sol: 61) - plus HLE 59.1% and Terminal-Bench v2.1 91.4%, both their best to date.

The cost is the catch. Artificial Analysis puts Fable 5.1 (max) at $3.76 per Intelligence Index task, 20% more than Fable 5 (max), driven by roughly 1.7x the output tokens - the cache-read cut does not offset it. Output speed is about 60 tok/s, slower than the frontier average. Anthropic also discloses that safety interventions zeroed some OSWorld and AutomationBench tasks, and Artificial Analysis observed roughly 4% of output tokens served via the fallback path to Opus models when the safety classifier intercepted dual-use requests.

Honest framing. The independent Intelligence Index win is the strongest signal; the vendor benchmark deltas and the safety-intervention disclosure should be read together, since agentic scores partly reflect where the classifier stepped in. The watermark is not priced in anywhere - it simply rides along on every output.