Claude Fable 5.1 vs Fable 5: 7 Things That Changed

On September 1, Anthropic rolled out Claude Fable 5.1 and claims it is up to 45% less costly to run than Fable 5, despite no changes in token pricing. The whole discount stems from billing for reused context, one of the lines in the pricing table named "cache reads."

The 45% figure is valid only for context-intensive agentic tasks. For everyday use cases, Anthropic gives its own estimate of 25%. The model launched alongside Claude Mythos 5.1, a restricted variant of the model built for vetted cybersecurity and biology researchers.

Fable 5.1 outperformed Fable 5 on Terminal-Bench-Science, Anthropic's test of AI-enabled scientific research, scoring 52.6% against 24.7%.

Price per million tokens for input and output stays the same as Fable 5, at $10 and $50. Anthropic's other model, Claude Opus 5, is priced at $5 and $25, and lags behind Fable 5.1 by three points on the Artificial Analysis Intelligence Index, a composite score across major AI benchmarks, 63 to 66.

You can make use of Fable 5.1 using usage credits on the Pro plan, just like Fable 5, and at no extra cost on the Max plan. 

/1. Fable 5.1's cache reads got 75% cheaper, and no other price moved

Cache reads are what Fable 5.1 charges each time it reuses context it has already processed: a system prompt, a tool result, a file- rather than processing that context again from scratch.

The rate dropped from $1 per million tokens under Fable 5 to $0.25 per million under Fable 5.1, a 75% cut. Anthropic attributes almost all of its up-to-45% savings claim to this change alone, specifically the work that relies heavily on context and tools.

If your task is a single short prompt with no context to reuse, you won't see much of a discount. For an average, mixed workload, it's closer to 25%. The model is only as cheap as the workload you use it for.

/2. The new model performed over twice as well on Anthropic's own scientific exam

On Terminal-Bench-Science, a benchmark that measures AI-driven scientific research, Fable 5.1 scored 52.6% compared to Fable 5's 24.7%. Opus 5 scored 29.0%, and OpenAI's GPT-5.6 Sol lagged behind all of them at 22.4%.

Anthropic also claims the model achieves a nearly 50% hit rate designing molecular binders, another task related to protein research. It also charted Venus's elevation map at a higher resolution than before.

Almost nobody building on the API will ever touch either feature. They're the clearest sign this version was built for research labs first.

/3. Fable 5.1 codes more effectively than Fable 5

Fable 5.1 scored 55.8% on Terminal-Bench 4.0, Anthropic's agentic coding benchmark, compared to 42.0% for Fable 5, 52.3% for Opus 5, and 37.3% for GPT-5.6 Sol. On CursorBench, a separate benchmark run through the Cursor editor, it climbed from Fable 5's 70.5% to 73.4%, narrowly surpassing Opus 5's 70.0%.

Fable 5.1 clearly outperforms Fable 5, Opus 5, and GPT-5.6 Sol on both coding benchmarks Anthropic published. Coding is the one category where the newer model wins outright, no caveats.

/4. The new model is okay at using a computer, but not great