Claude Fable 5.1 Keeps Its Price but Cuts Cache Reads by 75%

$10 per million input tokens, $50 per million output. That is the list price of Claude Fable 5.1, which Anthropic released on September 1, and it is exactly what Fable 5 cost. The one number that moved is cache reads: down from $1 to $0.25 per million tokens, a 75% cut. By Anthropic's own arithmetic that single change makes a typical workload about 25% cheaper and long-running agent jobs, which re-read the same context over and over, up to 45% cheaper.

The second half of the announcement is the more unusual one. Alongside Fable 5.1, Anthropic introduced Claude Mythos 5.1 and described the pair in one sentence: "the same model, but with different levels of safeguards." Same weights, different refusal policy.

What the benchmark deltas actually show

Across the five headline comparisons Anthropic published, the largest jump is in agentic scientific research: Terminal-Bench-Science goes from 24.7% on Fable 5 to 52.6% on Fable 5.1, more than double. Agentic coding on Terminal-Bench 4.0 rises from 42.0% to 55.8%, and computer use on OSWorld 2.0 from 72.9% to 77.9%. The gains on CursorBench and on Humanity's Last Exam are one to three points.

The shape of that table is the story. The improvement is concentrated in tasks that run for hours and plan their own intermediate steps, not in single-turn chat. Anthropic's showcase examples fit the same pattern: protein binders with 10 times the binding affinity of competing designs, a Venus elevation map sharpened from 10-20 km resolution to 2-3 km, and computational biology models tuned to run up to 2.5 times faster with 30-60% lower GPU cost.

Simon Willison's day-one notes add a figure developers should keep in mind. The model ships with five reasoning levels (low, medium, high, xhigh, max) and no option to switch reasoning off. The same prompt produced roughly 2,000 output tokens for $0.10 at "low" and 66,000 tokens for $3.30 at "max". A 33-fold spread means the effort setting, not the list price, will decide most bills.

Who gets Mythos 5.1

Fable 5.1 is available everywhere Claude is: Claude.ai, Claude Code, Claude Enterprise, and AWS, Google Cloud and Microsoft Azure, with the API identifier claude-fable-5-1. Mythos 5.1 is limited, for now, to US organizations admitted through two trusted-access tracks: a Cyber Verification Program for defensive security work and a Life Sciences Verification Program run with the US government.

The safeguards themselves changed in measurable ways. Anthropic says false refusals in cybersecurity contexts fell by 60%, biology safeguards fire 85% less often on benign medical and basic-biology requests, and vulnerability discovery is now permitted when the purpose is defensive. Outputs carry an invisible text watermark aligned with the EU AI Act's transparency requirements, and new anti-distillation mechanisms make it harder to clone the model through the API.

The cost math for teams already in production

Our reading: the cache price is the headline, not the benchmark table. Most production Claude deployments we see are either a customer-support assistant or a document-processing pipeline. In both, the prompt is largely fixed: system instructions, a knowledge base, worked examples. That fixed block is read from cache on every call, so the bill drops without touching a line of code.

A concrete scale: a support bot handling 200,000 calls a month, each reading 8,000 tokens of fixed context, consumes 1.6 billion cache tokens monthly. At the old price that was $1,600; at the new price it is $400.

Two caveats. A team that leaves reasoning at "high" and runs long tasks can hand the cache savings straight back in output tokens, so effort level is now a cost decision. And Mythos 5.1 access is limited to US organizations, which leaves security firms elsewhere without an application path for the unrestricted model at the moment.

Sources: Anthropic, Simon Willison, TechCrunch