Claude Fable 5.1 made me a nice animated pelican

On September 1st, 2026, a report by Simon Willison detailed the impressive capabilities of Claude Fable 5.1, an Anthropic model. The model demonstrated remarkable performance in coding, knowledge work, and long-running problem-solving tasks, achieving a new standard according to Anthropic's announcement. Among the various benchmarks, the Terminal-Bench-Science 0.1 showed the most impressive improvement, with Fable 5.1 scoring a remarkable 52.6% compared to previous models.

Willison focused on the pelican benchmark, a task used to evaluate how well the models could handle a specific prompt. While the benchmark's connection to overall model performance had become less clear, Willison found value in comparing model performance within the same family and at different reasoning effort levels.

He tested the model with various reasoning settings, from low to max. In low and medium settings, the model did not seem to execute reasoning at all, as the output token count remained relatively low and the reasoning text was missing. However, at higher reasoning levels, the model produced significantly more detailed reasoning traces, taking longer and costing more.

The "xhigh" setting produced an SVG of a pelican riding a bicycle, complete with intricate details like the bird's long neck, orange beak, and the bicycle's components. The "max" setting yielded the best result to date, with a pelican that was not only well-drawn but also included thoughtful additions like a blue hat, a basket with a fish, and a charming helmet.

Willison noted that while the result was impressive, it was not as elegant as Gemini 3.7 Flash's output. However, he maintained that the pelican was still a testament to the model's capabilities. The model's ability to create such detailed and well-designed SVGs, even at higher reasoning levels, was enough to make Willison question if the animated version could be created without incurring additional costs.