Anthropic today released Claude Fable 5.1 (and Mythos 5.1). The company says Fable 5.1 "sets a new standard for coding, knowledge work, and long-running problem-solving tasks." The announcement emphasizes scientific performance: the model scored 52.6% on the newly introduced Terminal-Bench-Science 0.1 benchmark (the benchmark was announced on August 27). For comparison, Fable 5 previously scored 24.7%, Opus 5 scored 29.0%, and GPT-5.6 Sol scored 22.4% on the same benchmark.
Other established benchmarks show modest improvements for Fable 5.1, but none as striking as the Terminal-Bench-Science 0.1 result.
Why the pelican SVG test matters
In a July piece, the author described losing faith in the pelican benchmark’s usefulness: its correlation with other model capabilities appeared weaker than it had been in 2025. The pelican test remains useful for within-family comparisons and for observing how a single prompt behaves at different reasoning effort levels.
Fable 5.1 offers five reasoning effort levels: low, medium, high, xhigh and max; there is no option to turn reasoning fully off.
The author fixed a bug in the llm-anthropic client that prevented reasoning traces from being recorded correctly, then re-ran a set of prompts. One examined prompt was: “Generate an SVG of a pelican riding a bicycle.” The author reproduced the full set of pelicans and the complete reasoning transcripts for each effort level.
Results by effort level
-
Low: the transcript did not show summarized reasoning tokens; the output token count was 1,998. The run took 23.8 seconds and cost 10.017 cents. (Note: with Claude, the output token count includes reasoning tokens.)
-
Medium: similarly showed no visible reasoning text; 1,977 output tokens, 23 seconds, 9.912 cents. This was 21 tokens fewer than the low run. For this prompt, Fable 5.1 appeared to skip producing visible reasoning at both low and medium settings.
-
High: this run produced a short reasoning segment. Output token count was 2,612, runtime 29.6 seconds, cost 13.087 cents. A brief summary from the reasoning read: “I'm planning the SVG layout for a pelican riding a bicycle, with a sky and ground background, a bicycle with two spoked wheels, frame, seat and handlebars, and a white-bodied pelican with a long neck and orange beak positioned on top.” Overall the difference from low/medium was modest.
-
XHigh (xhigh): at this level the output changed dramatically. Output tokens: 36,767; runtime: 7 minutes 51 seconds; cost: about $1.83. The reasoning trace was lengthy and detailed, including decisions like intentionally oversizing the pelican relative to the bike for comic effect and accepting slightly thicker strokes as charming rather than overengineering.
-
Max: the author judged this to be the best pelican they’d seen from any Anthropic model. Output tokens: 65,927; runtime: 13 minutes 54 seconds; cost: $3.30. The long reasoning trace contained many small design decisions (e.g., refining pedal placement relative to feet, adding a hat and a basket with a fish, resolving potential collisions between a helmet and the beak, adding darker feather tips, and choosing not to add unnecessary decorative elements). The author notes the result lacked the same degree of “flair” as Gemini 3.7 Flash, but it fulfilled the request for a precise SVG.
Quoted segments from the max reasoning trace illustrate the iterative vector-design thinking: adjusting control points on the front fork to correct rake, checking helmet vent placements relative to stroke widths and rounded caps, and reconsidering scalloped feather curves for a more natural trailing edge.
Animation follow-up
On Hacker News, user swalsh asked whether an animated version could be produced. Rather than pay another roughly $3, the author fed the Max SVG back into the model at the default High thinking level using the command llm logs -cx | llm -m claude-fable-5.1 -s 'animate this'. That run used 6,121 input tokens and produced 26,201 output tokens, costing $1.37. The exported animation shows the wheels turning in the wrong direction, but otherwise provides a pleasing motion derived from the original Max SVG.
Why this matters in practice
The Fable 5.1 announcement and the Terminal-Bench-Science 0.1 result suggest Anthropic is prioritizing scientific and long-horizon problem-solving capabilities. The pelican test highlights that reasoning effort settings can dramatically change output length, detail and cost: the highest settings produce large, human-readable planning traces and correspondingly high token usage and expense.
Practically, if a task requires precise, iterative design or planning (for example complex SVG generation), higher reasoning levels can yield better outputs but at substantially greater time and monetary cost. Conversely, low and medium levels are faster and cheaper but may omit visible reasoning, reducing transparency into the model’s internal decision process.
Conclusion
Claude Fable 5.1 shows notable gains on a targeted science benchmark and demonstrates flexible behavior across reasoning effort settings. For developers and artists using LLMs for structured generative tasks, the trade-offs between runtime, cost, and the visibility of stepwise reasoning remain important considerations.
