Claude Fable 5.1 generates an impressive animated pelican

Anthropic’s latest claim

Anthropic announced that Claude Fable 5.1 “sets a new standard for coding, knowledge work, and long‑running problem‑solving tasks”.

The company highlighted a 52.6 % score on the newly released Terminal‑Bench‑Science 0.1 benchmark.

That score improves on the 24.7 % recorded for Fable 5, the 29.0 % for Opus 5 and the 22.4 % for GPT‑5.6 Sol.

Other benchmark results also show modest gains, though none approach the science benchmark’s jump.

Testing the model’s creative side

On 1 September 2026 the author of the post ran a prompt asking the model to “Generate an SVG of a pelican riding a bicycle”.

Claude Fable 5.1 offers five distinct reasoning effort levels—low, medium, high, xhigh and max—and does not provide a way to disable reasoning entirely.

The author first patched a bug in the llm‑anthropic library that had previously prevented reasoning traces from being logged.

At the low effort setting the model produced 1,998 output tokens in 23.8 seconds, costing 10.017 cents, and the transcript contained no summarized reasoning tokens.

The medium setting yielded 1,977 output tokens in 23 seconds for a cost of 9.912 cents, again without any visible reasoning text.

These two settings therefore appeared to skip the reasoning phase for this particular SVG request.

When the effort was raised to high, the model generated 2,612 output tokens over 29.6 seconds at a cost of 13.087 cents, and a brief reasoning summary was recorded.

The high‑level reasoning note 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.”

Although the description added a few details, the visual result was not dramatically different from the low and medium outputs.

Increasing effort to xhigh produced a dramatic shift: 36,767 output tokens were emitted over 7 minutes 51 seconds, costing $1.83.

The extended reasoning trace included lines such as: “Adding the eye, wings stretching down to the handlebar grip, orange legs reaching to the pedals, and a small tail feather, while keeping the pelican intentionally oversized compared to the bike for comic effect.”

This level of detail shows the model investing significant token budget into planning the illustration.

At the maximum effort setting the model spent 65,927 output tokens across 13 minutes 54 seconds, costing $3.30, and produced what the author described as the best pelican generated by any Anthropic model to date.

The final SVG featured a tasteful background, legs clearly on either side of the frame, feet on the pedals, a wing on the handlebars, a cute blue hat on the pelican, and a basket containing a fish.

The author noted: “There’s a lot to like about this. The background is tasteful, the legs are clearly on either side of the frame, the feet are on the pedals, the wing is on the handlebars, the pelican has a cute blue hat and there’s a basket with a fish.”

While the output did not reach the flamboyance of Gemini 3.7 Flash, it fulfilled the explicit SVG request without unnecessary embellishment.

Implications of cost and token usage

The cost progression from under ten cents at low effort to over three dollars at max illustrates how reasoning depth directly impacts both compute time and expense.

Users seeking detailed visual assets may need to balance token consumption against budget constraints.

The experiment also demonstrates that Anthropic’s new reasoning controls can be leveraged to fine‑tune output quality for creative tasks.

Overall, the findings provide concrete evidence that higher reasoning settings yield richer visual descriptions at a predictable price.

Why This Matters: The data shows that Claude Fable 5.1’s tiered reasoning levels let users trade off cost for increasingly detailed SVG artwork, confirming Anthropic’s claim of improved problem‑solving capability.