Claude Fable 5.1 Hits 52.6% on Science Bench; Max Run Costs $3.30

Anthropic shipped Claude Fable 5.1 on September 1, 2026, positioning it as an advance for coding, knowledge work, and long-running problem-solving. The headline benchmark is a 52.6% score on Terminal-Bench-Science 0.1 — a benchmark that appeared on August 27th — compared to 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol. Other benchmarks showed slightly improved scores, but none as large as the science task gap. Mythos 5.1 was announced alongside Fable 5.1, though hands-on analysis from practitioner Simon Willison focuses exclusively on Fable.

Five Reasoning Tiers — No Off Switch

Fable 5.1 ships with five named reasoning levels — low, medium, high, xhigh, and max — with no option to disable reasoning entirely. Willison tested all five against a single SVG generation prompt ("Generate an SVG of a pelican riding a bicycle"), and the variance in cost and latency is significant.

At low, the model produced 1,998 output tokens in 23.8 seconds at 10.017 cents, with no visible reasoning trace. Medium was nearly identical: 1,977 output tokens, 23 seconds, 9.912 cents — and also no apparent reasoning. For this prompt, Fable 5.1 appeared to skip reasoning entirely at both settings.

High introduced a brief planning trace and consumed 2,612 output tokens over 29.6 seconds at 13.087 cents. The reasoning summary amounted to a layout sketch. The real inflection comes at xhigh: 36,767 output tokens, 7 minutes 51 seconds, $1.83. At max, the model consumed 65,927 output tokens over 13 minutes 54 seconds at $3.30, producing output with explicit deliberation over fork geometry, feather rendering, helmet-beak collision avoidance, and pedal placement.

Token Economics and Agentic Chaining

The cost curve between high and xhigh is not linear — output tokens jump roughly 14× and dollars follow proportionally. For teams routing tasks programmatically, selecting the wrong reasoning tier will dominate inference budgets quickly, which connects to the broader challenge of automating prompt and configuration optimisation in production pipelines.

Willison demonstrated a practical chaining pattern: he piped the max-effort SVG output directly into a fresh high-effort call to animate it, using 6,121 input tokens and 26,201 output tokens at a combined cost of $1.37. The total for the full max-then-animate workflow was $4.67. The animated result had wheels rotating — though Willison notes the apparent reverse rotation is likely an MP4 conversion artifact rather than a model error.

Benchmark Context

The Terminal-Bench-Science 0.1 score deserves scrutiny because the benchmark is new: Anthropic is the primary beneficiary of a benchmark announced four days before the model launch, making independent replication important before treating the 52.6% figure as settled. The internal family comparison is informative on its own terms — Fable 5.1 more than doubles Fable 5's 24.7% and substantially clears Opus 5's 29.0% — but the magnitude of that gap is precisely why external validation matters.

The broader signal is that Anthropic is betting on tuneable reasoning depth as a product differentiator. The five-tier structure forces developers into explicit cost-quality tradeoffs rather than treating inference as a fixed-cost commodity — a design philosophy with direct consequences for how teams architect agentic pipelines where model selection alone cannot guarantee safe or efficient outcomes. Whether the Terminal-Bench-Science advantage survives external scrutiny will determine how seriously competitors respond.