Anthropic Releases Claude Fable 5.1, Claims 52.6% on New Terminal-Bench-Science Benchmark

Anthropic Ships Claude Fable 5.1

Anthropic released Claude Fable (and Mythos) 5.1 today, claiming the model "sets a new standard for coding, knowledge work, and long-running problem-solving tasks." The company's announcement emphasizes scientific research performance, reporting a 52.6% score on Terminal-Bench-Science 0.1 — a benchmark first published on August 27, 2026.

That figure represents a substantial jump from Fable 5's 24.7% on the same benchmark. According to Anthropic, Fable 5.1 also outperforms competing models: Opus 5 scored 29.0% and GPT-5.6 Sol scored 22.4% on Terminal-Bench-Science 0.1. Anthropic describes gains on other benchmarks as more modest.

Five Reasoning Levels, No Off Switch

Fable 5.1 introduces five reasoning effort settings — low, medium, high, xhigh, and max — with no option to disable reasoning entirely. Developer and independent tester Simon Willison ran identical prompts ("Generate an SVG of a pelican riding a bicycle") across all five levels after fixing a bug in the llm-anthropic library that had prevented reasoning traces from being logged correctly.

The results show enormous variance in cost and latency for the same task:

• Low: 1,998 output tokens, 23.8 seconds, $0.10
• Medium: 1,977 output tokens, 23 seconds, $0.099
• High: 2,612 output tokens, 29.6 seconds, $0.13
• Xhigh: 36,767 output tokens, 7 minutes 51 seconds, $1.83
• Max: 65,927 output tokens, 13 minutes 54 seconds, $3.30

At low and medium settings, the model appeared to skip reasoning almost entirely — no summarized reasoning tokens appeared in either transcript, and medium actually used 21 fewer output tokens than low. High effort produced a brief reasoning summary but a nearly identical image. The jump to xhigh and max produced substantially longer reasoning traces and visibly more detailed SVG output, including a version with a hat and a fish basket at the max setting.

Based on the cost figures, Fable 5.1 pricing works out to approximately $10 per million input tokens and $50 per million output tokens — in line with Anthropic's premium-tier pricing for prior Opus-class models.

What This Means

The headline benchmark claim — more than doubling Terminal-Bench-Science performance versus Fable 5 — is notable but comes from a benchmark that is less than a week old and has no independent track record. Buyers should treat the 52.6% figure as an Anthropic claim until third-party evaluations confirm it.

More practically significant for developers: the five-tier reasoning system creates a 33x cost spread ($0.10 to $3.30) and a 35x latency spread (23 seconds to nearly 14 minutes) for the same prompt, with no way to turn reasoning off completely. That's a meaningfully different cost model than binary reasoning toggles used by competitors, and it means teams building on Fable 5.1 need to actively tune the effort parameter per use case rather than treating reasoning as an on/off switch. Low and medium settings appear functionally similar for simple generation tasks, which suggests the medium tier may be redundant for many use cases — testing on your own workload before committing to a default effort level will matter more than usual.