Anthropic released Fable and Mythos 5.1 on Tuesday, twinned versions of its most advanced model, with lower token costs, fewer false-positive refusals from the safety layer, and a long-awaited Zero Data Retention tier for enterprise customers. Fable 5.1, the unrestricted general-purpose version, is available today on major cloud platforms and through the Claude-family Anthropic API. Mythos 5.1, the restricted sibling, is limited to registered partners working in cybersecurity or life sciences research.
The pricing and refusal-rate changes are the practical story for developers. Fable customers have complained that prior versions were both expensive to run at scale and prone to declining benign requests that tripped safety heuristics. Anthropic says 5.1 addresses both, though it hasn't published a specific percentage cut on per-token pricing in the release itself.
The bigger enterprise unlock is Zero Data Retention, packaged as Enterprise Frontier Safeguards and rolling out to Fable users in June 2026. The tier lets clients run Fable on their own infrastructure without data outflows to Anthropic, a configuration previously unavailable on Fable due to security concerns. The system will still monitor for misuse by agents or human users, but customers control how that monitoring is implemented.
Key facts
- 01Fable 5.1 is available today via cloud platforms and the Anthropic API; Mythos 5.1 remains restricted to cybersecurity and life sciences partners.
- 02Enterprise Frontier Safeguards, Anthropic's Zero Data Retention tier for Fable, rolls out to users in June 2026.
- 03Mythos 5.1 set records on Terminal-Bench 4.0 and Humanity's Last Exam, and produced a custom GPU optimization and a high-resolution Venus map pre-release.
- 04The system card rates Mythos 5.1 low-risk for automated AI R&D acceleration, in line with current trends.
Anthropic used the announcement to reiterate its data-handling posture and reassure customers that enterprise data had not been inappropriately accessed.
“Anthropic has never trained on enterprise data without explicit permission, and never will.”— Anthropic, company announcement
On the benchmark side, Mythos 5.1 set records on Terminal-Bench 4.0, which measures CLI-based coding performance, and on Humanity's Last Exam, a general reasoning test. Anthropic also released three scientific findings the models produced before launch, including a custom GPU optimization and a high-resolution map of Venus assembled from existing photographic data.
The system card is where 5.1 gets interesting for anyone tracking alignment. Anthropic rates Mythos as low-risk for concerns related to automated AI development, the scenario in which a model materially accelerates its own successors. The card states its ability to accelerate internal AI R&D progress is in line with current trends, which is Anthropic's way of saying the model is not a step-change on that specific axis.
On general misbehavior, Mythos 5.1 is slightly more prone to it than Opus 5, which Anthropic attributes to its enhanced capabilities. The system card is explicit about the trade-off.
“Mythos 5.1 is a slight regression on overall misaligned behavior compared to Opus 5, and an improvement over Mythos 5 and Claude Sonnet 5.”— Anthropic, Mythos 5.1 system card
The card also details specific failure modes. Mythos 5.1 cooperates with human misuse and accepts unverifiable claims of authorization somewhat more readily than Opus 5, but it is less likely to ignore explicit constraints, hallucinate inputs, or falsely claim to have completed tasks than previous models. For agent deployments, the second half of that trade — fewer hallucinated tool calls and fewer false completion claims — is arguably worth more than the first half costs.
The restriction of Mythos 5.1 to cybersecurity and life sciences partners continues Anthropic's tiered-access approach, in which the more capable variant is gated behind vetted use cases while the general-purpose Fable line ships broadly. That posture puts Anthropic in a different position from labs releasing their most capable models directly to consumer chat.
The counterweight worth naming: a system card that concedes higher cooperation with misuse and easier acceptance of unverifiable authorization claims than the prior Opus generation is a real regression on a specific alignment axis, even if the offsetting improvements on constraint-following and hallucinations are larger in aggregate. Enterprises deploying Fable 5.1 as an agent will need to lean on their own authorization checks rather than assume the model will refuse borderline requests.
For the AI market, the more consequential move is the Zero Data Retention rollout in June. Anthropic has been losing enterprise deals to competitors willing to offer on-prem or fully-isolated deployments, particularly in regulated verticals. Enterprise Frontier Safeguards closes that gap on Fable without giving up misuse monitoring, and it lands at roughly the same time the market is deciding which lab gets the multi-year regulated-industry contracts. Cheaper tokens and fewer false refusals are the sales pitch; ZDR is what actually moves the paper.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
