Claude Fable 5.1 claims 52.6% on new science benchmark; pelican test runs 10 cents to $3.30

Anthropic pitches Fable 5.1 on science benchmarks

Anthropic released Claude Fable 5.1 on 1 September 2026, alongside Mythos 5.1. According to Simon Willison, who tested the model on launch day, Anthropic presents it as setting a new bar for coding, knowledge work and long-running problem-solving, and the announcement leans noticeably on scientific research.

The headline number is 52.6% on Terminal-Bench-Science 0.1, a benchmark first announced on 27 August. As Willison relays the comparisons, Fable 5 scored 24.7%, Opus 5 scored 29.0% and GPT-5.6 Sol scored 22.4%. Other benchmarks in the announcement improved only marginally, which makes the science result stand out even more.

The pelican test

Willison's informal counterweight to vendor benchmarks is a prompt he has run for years: ask a model to generate an SVG of a pelican riding a bicycle, then look at what comes back. He wrote in July that he is losing faith in it as a general capability proxy, because its correlation with other skills has weakened since 2025 — but he still finds it useful for comparing models within a family, and above all for comparing reasoning effort levels on the same prompt.

That matters here because Fable 5.1 exposes five reasoning levels — low, medium, high, xhigh and max — with no way to turn reasoning off entirely. After fixing a bug in his llm-anthropic plugin that had stopped reasoning traces from being recorded correctly, Willison ran the pelican prompt at every level.

Ten cents to $3.30 on one prompt

The spread between levels is dramatic. At low effort the model produced 1,998 output tokens in 23.8 seconds for 10.017 cents, with no reasoning text visible. Medium was oddly similar: 1,977 tokens — 21 fewer than low — in 23 seconds for 9.912 cents, and also no reasoning shown, a puzzle Willison flags given that Claude output token counts normally include reasoning tokens. High added a brief planning summary, 2,612 tokens, 29.6 seconds and 13.087 cents, with output little different from the cheaper tiers.

The step change came at xhigh: 36,767 output tokens, 7 minutes 51 seconds and $1.83. The reasoning trace shows genuine design deliberation, including a decision to keep the pelican deliberately oversized next to the bike for comic effect. At max — 65,927 tokens, 13 minutes 54 seconds, $3.30 — Willison calls the result the best pelican he has seen from any Anthropic model: a tasteful background, legs on either side of the frame, feet on the pedals, a wing on the handlebars, a blue hat and a basket carrying a fish.

The max reasoning trace reads like a designer's notes. The model weighed a bicycle helmet against the pelican's signature crest, shrank the helmet so it would not collide with the beak, checked that helmet vents stayed inside the helmet's outline, decided against a handlebar bell as unnecessary, and moved a control point so the front fork bowed forward with proper rake. Willison still rates Gemini 3.7 Flash higher on flair, though he notes he asked for an SVG, and that is what he got.

Then he animated it

When a Hacker News commenter asked for an animated version, Willison piped the max pelican back into Fable 5.1 at its default high effort with the instruction to animate it. That run cost $1.37 on 6,121 input and 26,201 output tokens. Exported as video, the wheels appear to rotate backwards, but he believes that is an artifact of the MP4 conversion — in the original SVG they spin the correct way.

Why it matters

Two things stand out. First, the launch's most impressive figure rests on a benchmark that was five days old at release, and where the gap to predecessor models is unusually large. Fresh benchmarks with little public track record can dominate a launch narrative, which is why independent checks like Willison's carry weight.

Second, the pelican run is an unusually clear picture of the cost-quality dial on modern reasoning models: the identical prompt cost about 10 cents at low effort and $3.30 at max — roughly a thirty-fold spread — with the visible quality gains concentrated in the top two tiers. Choosing a reasoning level is now a budget decision as much as a quality one, and since reasoning cannot be disabled, the API, not the developer, sets the floor.