Claude Fable and Mythos 5.1: one model, two permission tiers

On Tuesday, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, twin versions of its most advanced model, according to TechCrunch: the same underlying model, with different levels of guardrails. Fable 5.1, the unrestricted version, became available on major cloud platforms and through Anthropic's API as soon as it was announced.

Mythos 5.1, by contrast, follows the same approach as the previous Mythos generation: it is available only to Anthropic's registered partners working in cybersecurity or life-sciences research, not to the public. This "same model, different permissions" distribution means a company must first demonstrate a verified use case before it can access the deeper capability boundary.

Shared intelligence, tiered permissions

According to Leiphone's technical analysis, the real change in this release is not the model's level of intelligence but how Anthropic now fully separates model capability from the scope of execution permissions: Fable handles general coding, knowledge work and everyday agent tasks; Mythos targets verified cybersecurity and life-sciences scenarios, with access to deeper tool calls and more specialized tasks. In other words, the two versions share the same "brain", but not the same reach.

TechCrunch reports that the new models set records on several benchmarks, including Terminal-Bench 4.0, which measures command-line coding, and Humanity's Last Exam, which tests general reasoning. Anthropic also published three previously unpublished scientific discoveries generated by the models before release, including a bespoke GPU kernel optimization and a high-resolution map of Venus assembled from existing photographs.

• Fable 5.1: unrestricted version, available on cloud platforms and via the Anthropic API as soon as it was announced
• Mythos 5.1: limited to registered cybersecurity and life-sciences partners, as with the previous generation
• New records on Terminal-Bench 4.0 (command-line coding) and Humanity's Last Exam (general reasoning)
• The high-confidentiality "Enterprise Frontier Safeguards" service will extend to Fable this fall, letting customers keep data on their own infrastructure

38 hours without sleep, and a change of mind along the way

Leiphone cites a test run by the fintech company Ramp: Fable 5.1 ran unsupervised for 38 hours. During that time, it identified on its own a mislabeled artifact in the running experiment, reprocessed the affected data, launched six experiment groups in parallel, and kept adjusting its original plan as results came in.

The significance, the analysis argues, is not the duration itself but the fact that the model revised an already-established judgment when experimental conditions changed, rather than mechanically following its initial plan. That points to a deeper problem for long-running agents: over time, code may have been modified several times, data versions change, and old conclusions get invalidated by new results while still sitting in the conversation history — what Leiphone calls "state drift". The article argues that simply widening the context window is not enough; what's needed is an explicit mechanism for invalidating stale information, or a long-running agent ends up "clinging to a pile of outdated history and confidently rambling".

Three scientific discoveries, one working pattern

Leiphone examines several research tasks Anthropic showcased — protein design, reconstructing the terrain of Venus, GPU kernel optimization — and finds they share the same structure: Claude does not directly run the specialized computation itself, it orchestrates multiple specialized tools within a continuous workflow. That requires asynchronous execution: the model submits a run, keeps the job ID, moves on to other parallel tasks, then reconnects once a result comes back. Ramp's six parallel experiments are a direct example of this.

This way of working also requires "provenance": every intermediate result has to be tied to the exact version of the data, model, code and parameters used to produce it, or incompatible results end up mixed together and lead to wrong conclusions.

System card: "low" risk, but also a slight regression

According to TechCrunch, Anthropic's system card rates Mythos's risk on automated AI research and development — AI improving AI — as "low", a topic some see as a possible trigger for loss of human control. On general misaligned behavior, Mythos is slightly more prone to it than Opus, possibly because of its enhanced capabilities.

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. It 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.

TechCrunch highlights a notable change in this release: broader adoption of zero data retention. Enterprise customers can now run Anthropic's models on their own infrastructure without data ever leaving it. This high-confidentiality service, called "Enterprise Frontier Safeguards", had until now been unavailable for Fable for security reasons; it will roll out this fall. Anthropic says it will still monitor misuse by agents or humans, but customers will control how that monitoring is carried out. The company also stresses that customer data has never been inappropriately accessed: "Anthropic has never trained on enterprise data without explicit permission, and never will."

Leiphone also flags an evaluation pitfall that's easy to overlook: model-level results and agent-level results can no longer be conflated. A failed task might stem from a bad model judgment, an unaddressed stale state, a failed tool call, a permission-system error, or a parallel-scheduling problem. That calls for evaluation that examines the full execution trajectory, not just the final success rate.

For developers and companies rolling out autonomous AI agents in production, this split between model intelligence and execution permissions offers a concrete template for grading risk: rather than waiting for a more capable model to become automatically safer, teams can grade scenarios upfront — routine work handled by a broadly permissioned version, higher-risk cases such as security testing or biomedical research routed through a verified channel with deeper access and tighter oversight. As agents capable of running for dozens of hours become more common, questions of state drift and result provenance will land on the desk of every team building long-running task systems.