Anthropic Releases Claude Fable 5.1 with Major Breakthrough in Scientific Reasoning
Anthropic has launched Claude Fable 5.1, reaching a 52.6% score on the new Terminal-Bench-Science 0.1 benchmark, more than double the 24.7% score achieved by Fable 5. Tech analyst Simon Willison evaluated the release using his custom SVG pelican benchmark across the model's five configurable reasoning levels. The significant leap in scientific benchmark performance demonstrates rapid advances in handling complex domain-specific research workflows. In addition, qualitative testing of reasoning effort levels offers useful practical insight into how LLMs manage computational costs and output quality for creative coding tasks. Claude Fable 5.1 offers five explicit reasoning effort tiers (low, medium, high, xhigh, max) with no setting to turn off reasoning entirely. In Willison's initial pelican SVG test, the low and medium settings returned nearly identical token counts and skipped visual reasoning steps altogether.
## BACKGROUND
Simon Willison's pelican benchmark prompts LLMs to generate raw SVG code of a pelican riding a bicycle to qualitatively measure spatial awareness and graphic code synthesis. Terminal-Bench-Science is an evaluation framework designed by domain experts to test autonomous AI agents on scientific research workflows across disciplines like chemistry, physics, and biology.