Claude Fable 5.1 made me a nice animated pelican

What happened

On September 1, 2026, Anthropic shipped Claude Fable 5.1 alongside Claude Mythos 5.1, and the developer community wasted no time stress-testing the release. Simon Willison, writing on his weblog, ran one of the most widely followed informal evaluations in AI circles — the pelican benchmark, a repeatable SVG generation task that surfaces differences in reasoning depth and output quality across model families. The results reveal a model that scales dramatically with reasoning effort, but with costs and latency that make that ceiling impractical for most everyday tasks.

This article breaks down what shipped, how Fable 5.1's five reasoning levels behave in practice, what developers need to know before integrating the model, and where to watch for Anthropic's next moves. It is written for engineers and product teams already comfortable with calling the Anthropic API and familiar with the llm-anthropic Python plugin.

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. Anthropic's announcement positions Fable 5.1 as setting a new standard for coding, knowledge work, and long-running problem-solving tasks. The headline benchmark number is 52.6% on Terminal-Bench-Science 0.1, a new benchmark first announced on August 27, 2026. That figure compares favorably to 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol. Outside of the science benchmark, other scores show only modest gains over the previous release.

How it works

Fable 5.1 introduces five named reasoning effort levels: low, medium, high, xhigh, and max. There is no option to disable reasoning entirely. This is a departure from how earlier Claude models handled inference: every call now passes through at least some reasoning machinery, though at lower effort levels the model may elect not to surface any reasoning tokens at all. The Mythos 5.1 model was released simultaneously but Willison's testing focused on Fable 5.1.

The reasoning effort selection changes how you pay, how long you wait, and what you get back. At low effort, a single SVG generation prompt with 27 input tokens produced 1,998 output tokens in 23.8 seconds at a cost of 10.017 cents. Medium produced 1,977 output tokens in 23 seconds at 9.912 cents — marginally cheaper and faster than low, with no visible reasoning trace in either case. High added a short internal plan and produced 2,612 output tokens in 29.6 seconds at 13.087 cents.

Why it matters

Xhigh changes the character of the response entirely: 36,767 output tokens, 7 minutes and 51 seconds, and $1.83 for the same prompt. Max pushed to 65,927 output tokens, 13 minutes and 54 seconds, and $3.30. Because output token counts include reasoning tokens in Anthropic's billing, the xhigh-to-max jump is not just a quality dial — it is a cost multiplier of roughly 18x versus high effort. Builders should treat xhigh and max as deliberate choices for complex, long-horizon tasks, not defaults.

Willison accessed Fable 5.1 through the llm-anthropic plugin, specifying the model as claude-fable-5.1 and passing effort levels as parameters. Before running his pelican tests he fixed a bug in llm-anthropic (tracked as issue 88 on GitHub) that caused reasoning traces to not be correctly recorded. Any developer using the plugin to capture reasoning output should verify they are running a version that includes that fix before relying on trace data for debugging or evaluation.

Setting effort level in llm-anthropic follows the existing reasoning parameter convention for Anthropic models. Piping prior conversation context into a new call is supported through the llm logs -cx pattern, which Willison used to feed the max-effort SVG into a follow-up high-effort animation prompt. The animation call consumed 6,121 input tokens and 26,201 output tokens, costing $1.37. The full model identifier to target in API calls or llm invocations is claude-fable-5.1.

Who is affected

At low and medium effort, Fable 5.1 may produce zero visible reasoning tokens even though reasoning is nominally active. Willison observed that both low and medium runs of the pelican prompt showed no summarized reasoning text, and the token counts between the two levels were nearly identical — 1,998 versus 1,977. This means effort level is not a guarantee of reasoning depth; the model decides internally whether to reason for a given prompt, and simpler tasks may bypass it entirely regardless of the setting you pass.

Animated SVG output may render differently depending on the environment. The animated pelican generated from the max SVG played wheel rotation in what appeared to be the wrong direction when exported to MP4 at 720 by 540 pixels and 15 frames per second with a CRF of 30, though the original SVG appeared to animate correctly. If your pipeline converts SVG animations to video for delivery, verify rotation and timing in the source SVG before encoding, not only in the video output.

What to watch next

Willison noted that Gemini 3.7 Flash currently produces more visually expressive SVG output than Fable 5.1 at max effort, though the prompt he used asked only for a correct SVG and not for stylistic flair. That distinction matters: Fable 5.1 at max appears to closely follow literal instructions rather than embellish them. Watching how the model behaves on open-ended creative prompts versus tightly specified ones will clarify where in a production pipeline it earns its cost premium.

Terminal-Bench-Science 0.1 was announced just four days before the Fable 5.1 release, and Anthropic's announcement leaned heavily on its 52.6% result. It is worth monitoring whether that benchmark gains independent adoption or whether it was designed to highlight Fable 5.1's strengths. Willison's own pelican benchmark has shown diminishing signal strength for cross-family comparisons since 2025, though it continues to expose within-family differences clearly — including the sharp quality jump Fable 5.1 exhibits between high and xhigh reasoning.