Claude Fable 5.1

Anthropic's claude-fable-5-1, successor to Fable 5 in the same tier at the same per-token price: $10 / $50 per million input / output tokens, 1M context, 128K max output. Cache reads are $0.25/MTok. It is a Covered Model — zero-data-retention orgs are refused unless Anthropic authorizes them. It is not offered on Priority Tier.

Two API facts matter for reading benchmarks of it. Thinking is always on and cannot be disabled; depth is set through output_config.effort, which takes low, medium, high, xhigh and max, with high the default. And forced tool use is gone: tool_choice any and tool return a 400, so a harness that pinned the model to a tool has to move to auto plus an instruction, strict: true, or structured outputs. Thinking blocks are also bound to the producing model, and editing earlier turns invalidates them.

What the corpus measures so far

One third-party read, and it is narrow: coderabbit-fable-5-1-review (T2, 45 code-review tasks against 105 known issue points). On that harness Fable 5.1 scored 61.0% recall / 37.3% precision, against Fable 5's 61.9% / 32.8%. The same share of real defects found, with 34% fewer comments and 70% fewer nitpicks, for 49% more latency. Against Opus 5 (55.2% / 39.3%) and GPT-5.6 Sol (69.7% / 31.6%) it sits in the middle of both axes and wins neither.

The effort knob went the wrong way there. High reasoning was slower (21:36 vs 18:38) and found less (57.1% recall vs 61.0%). One harness, one task family, no repeat runs, but it is the first measurement in this spoke where raising effort cost time and returned nothing, and it is worth watching for a second instance before it means anything.

Nothing here covers this model's price/performance against its own predecessor, or any independent benchmark. The claude-fable-5 page's open question — whether the Fable tier's price is justified by capability — carries forward unanswered to 5.1.

One long-horizon data point arrived, and it is vendor-shaped. claude-nine-loop-amplitude-von-hippel (25 Sep 2026) has Fable 5.1, inside Anthropic's Claude Science harness, working largely unsupervised for hours-to-days and writing its own Python/SymPy code to reach a nine-loop amplitude in planar N=4 super-Yang-Mills — one loop past the best published human result (Dixon & Liu 2023, eight loops) — by two independent methods. It is the first source here on the agentic/long-horizon behaviour the Fable tier is sold for. It is also T3: Anthropic chose the problem, ran the harness, and paid the writer; the one outside validator (Lance Dixon) received Claude credits for checking it. That validation is real independent expertise, not a rubber stamp, but it is not a disinterested benchmark either — read it as one vendor-adjacent data point on autonomy, not a settled capability score.

Related

claude-fable-5 · claude-opus-5 · coderabbit-fable-5-1-review · claude-nine-loop-amplitude-von-hippel · llm-benchmarks · llm-api-pricing · anthropic