Anthropic released Claude Fable 5.1 and Mythos 5.1 on 1 September, and the number that travelled was “around 25% cheaper”. The pricing table on the announcement says something narrower: $10 per million input tokens and $50 per million output tokens, both unchanged from Fable 5. What changed is cache reads, down 75% to $0.25 per million.
What the common framing gets wrong
A discount that lives entirely in cache reads only exists for workloads that actually reuse a cached prefix. A one-shot call with no cache hit costs exactly what it cost last week. The “up to approximately 45%” figure quoted for agentic workloads is the best case, not the typical one, and it is best precisely because long-running agents re-read the same context repeatedly. This is not a price cut on intelligence; it is a price cut on repetition.
Two names, one model
The second thing coverage flattens is that Fable 5.1 and Mythos 5.1 are the same underlying model at different safeguard levels, with Mythos gated to US entities in trusted-access programmes. On Terminal-Bench 4.0 the pair score 55.8% and 60.9%, against Mythos 5 at 42.0%, Opus 5 at 52.3% and GPT-5.6 Sol at 37.3%. Read correctly, the 5.1-point gap between the two is not a capability difference at all — it is what the same weights do with fewer refusals. It is one of the few public numbers that prices a safety layer.
The benchmark with the error bar nobody quotes
The flagship result is Terminal-Bench-Science 0.1, where Fable 5.1 reaches 52.6% against Fable 5's 24.7%, Opus 5's 29.0% and GPT-5.6 Sol's 22.4%. Anthropic prints a stated error bar of ±3.5 to 4.5 points per model beside it. Elsewhere: GDPval-AA v2 at 1853 (Fable 5: 1723), OSWorld 2.0 at 77.9% partial and 41.7% strict, Humanity's Last Exam at 60.9% without tools, and AutomationBench at 31.4%.
Fewer interventions, and a narrower cyber rule
Anthropic says biology safeguards now fire 85% less often on benign requests and cybersecurity safeguards produce around 60% fewer interventions per session. The policy line moved with them: Fable 5.1 may be used to discover vulnerabilities, but not to develop exploits for them.