Claude Fable 5.1: Leap into Science and Modular Reasoning

Anthropic presents Claude Fable 5.1 with a score of 52.6% on Terminal-Bench-Science 0.1, significantly higher than the compared versions and models. Five levels of reasoning are offered, with no option to completely disable them. Tests show marked differences in effort and output, culminating in the generation of an SVG of a "bicycle pelican," followed by an animated version.

An animated "bicycle pelican" produced after detailed rendering

A generation performed at the maximum level resulted in a richly detailed SVG, incorporating a background, the pelican's two legs positioned on either side of the frame, its feet on the pedals, one wing on the handlebars, a blue hat, and a basket containing a fish. Following a suggestion from user swalsh on Hacker News, this version was used as a basis to generate an animation using the default reasoning level "High" via the indicated command. The operation utilized 6,121 tokens in input and 26,201 in output, costing $1.37. During the video export, the wheels appear to turn in the wrong direction, attributed to the conversion to MP4, while they appear correct in the original SVG.

Fable 5.1 introduces five reasoning levels without total deactivation

Fable 5.1 offers five reasoning levels—low, medium, high, very high, and max—without the option to completely disable this mechanism. A fix has been applied to the llm-anthropic tool to correctly record traces before the tests. In a test with the prompt "Generate an SVG of a pelican riding a bicycle," the low and medium settings did not show summarized reasoning text, with 1,998 tokens (23.8 s, 10.017 cents) and 1,977 tokens (23 s, 9.912 cents) respectively. The high level produced 2,612 tokens in 29.6 s (13.087 cents) with a brief outline of the SVG construction. The very high level generated 36,767 tokens in 7 min 51 s ($1.83) with a trace detailing proportions and elements, even accepting a "slight thickness" deemed charming. These observations show a gradual increase in effort and verbosity of traces according to the chosen level.

The debate around the pelican benchmark and the interest in internal comparisons

Concerns were raised in July regarding the correlation of the pelican benchmark with the overall quality of models, considered less obvious than in 2025. The reported interest mainly focuses on comparisons within the same family and between effort levels for the same prompt. Trace excerpts illustrate this granularity: adding pedals, hesitating over accessories, managing visual clutter, deciding between a helmet and a crest, and then making geometric adjustments such as reducing a helmet to avoid overlap at point (484.84), adding an area of dark feathers, checking the helmet's vents, and correcting the fork curve. For perception, Gemini 3.7 Flash is cited as having more "flair," although this was not an objective of the prompt.

Anthropic emphasizes science with a 52.6% on the new index

Anthropic announces Claude Fable 5.1 and Mythos 5.1, claiming that Fable 5.1 sets a new standard for programming, knowledge work, and long problem-solving tasks. The company boasts a score of 52.6% on the brand new Terminal-Bench-Science 0.1, presented for the first time on August 27. Published comparisons show Fable 5 at 24.7%, Opus 5 at 29.0%, and GPT-5.6 Sol at 22.4% on the same test. Other evaluations are mentioned with more modest progressions, less striking than the performance observed on the scientific component.