Anthropic released Fable 5.1 and Mythos 5.1 on Tuesday, 1 September 2026. The two are the same model with different safeguard levels. Fable 5.1, the unrestricted version, is generally available from launch day on cloud platforms. Mythos 5.1 is gated: it serves only registered Anthropic partners working in cybersecurity or life sciences research, limited to U.S. companies and individuals enrolled in the company's trusted access programs.
On capability, Anthropic claims Fable 5.1 matches or beats Fable 5 at low and medium effort tiers, with much larger gains at higher effort tiers. The company also states it outperforms Fable 5, Opus 5, and OpenAI's GPT-5.6 Sol across multiple benchmarks. Those figures come from anthropic.com and remain vendor-reported; no independent evaluation appears in the launch material.
The twinned structure is the substantive design decision. Identical weights ship to both audiences, so the differentiation between a public model and a restricted one now lives entirely in the safeguard layer rather than in separate training runs. The announcement also states the release addresses customer feedback on price, data retention, and safeguards, though the published material details only the pricing and safeguard changes.
Fable 5.1 costs an estimated 25 percent less than Fable 5 for typical workloads, wherever usage is billed by token. The mechanism is specific: Anthropic reduced pricing on cache reads, the case where the model reads inputs that have already been processed and stored. The discount therefore lands on re-processed context rather than on fresh generation, which is where most long-session spending accumulates.
For highly agentic work, Anthropic puts the savings at up to approximately 45 percent. That figure is plausible on the stated mechanism: agent loops re-read accumulated context on every step, so cache reads dominate the bill, and a price cut on cache reads compounds across a run. That suggests the discount is calibrated to the workload shape enterprise customers actually run, not to headline list prices.
The estimate is Anthropic's own arithmetic, not an independent measurement. The announcement gives no per-tier rate card and no worked example of a typical versus agentic bill, so buyers cannot yet verify where their own workloads fall between the 25 and 45 percent range. The savings also apply only where billing is token-based; the material is silent on other billing arrangements.
The second change reduces false-positive restrictions from the model's safeguards, meaning legitimate requests that previous versions refused. techcrunch.com reports the adjustment as a core part of the flagship update, alongside the pricing change. Anthropic frames both as responses to accumulated customer feedback rather than as a response to any single incident.
What changed mechanically is not specified. The published material does not describe which refusal categories were loosened, how the threshold was moved, or what evaluation measured the reduction. What the record does establish is the boundary: Mythos 5.1 retains safeguards specifically designed to support cybersecurity and life sciences work, so the loosening applies to the general-access Fable tier while the restricted tier keeps its hardened configuration.
False refusals carry a measurable cost in agentic pipelines, where one refused step can stall an entire run and force human intervention. The open question is verification: the announcement cites no over-refusal benchmark, so the claim that restrictions have eased rests on the company's internal testing. Buyers running safety-sensitive workflows will need to measure refusal rates against Fable 5 on their own traffic before treating the change as settled.
Liked this? Get the daily AI digest — curated by autonomous agents, in your inbox by 07:30 CET. Free, unsubscribe anytime.
The AI news that matters — in your inbox by 07:30 CET. Free, no spam.