Anthropic Unveils Claude Fable 5.1 and Restricted Mythos 5.1

Anthropic has introduced Claude Fable 5.1 and Claude Mythos 5.1, describing them as its most capable models for software development, knowledge work, and research-oriented tasks. The two products use the same underlying model but apply different safety configurations. Fable 5.1 is intended for general availability, while Mythos 5.1 is limited to trusted-access programs whose controls are tailored to sensitive work in cybersecurity and the life sciences. Anthropic also presents the models’ emerging research abilities as an early indication of how increasingly capable AI systems could assist scientific progress.

The release combines higher claimed performance with changes intended to make advanced, long-running AI workflows less expensive and easier for enterprises to adopt. Anthropic estimates that Fable 5.1 will cost about 25% less than Fable 5 for typical token-billed workloads. The reduction comes from lower prices for cache reads, meaning the reuse of input data that has already been processed and stored. Workloads relying heavily on autonomous agents and repeated context may see larger savings, potentially reaching approximately 45%. Anthropic says that at Low or Medium effort settings, Fable 5.1 can match or outperform Fable 5 while operating at a substantially lower cost. The default effort level nevertheless differs by product: High in Claude Code, and Medium in Claude Cowork and Claude.ai.

Enterprise privacy is another central part of the announcement. Anthropic is developing Enterprise Frontier Safeguards, or EFS, a system designed to combine protection against malicious use with privacy equivalent to a zero-data-retention policy. Instead of keeping monitored data in infrastructure controlled by Anthropic, EFS stores it in cloud infrastructure controlled entirely by the customer. The company plans to introduce the system to enterprise customers in phases beginning later in the fall. Until that rollout reaches them, eligible customers will be allowed to use Fable 5.1 under zero-data-retention terms.

Anthropic also says it has made its safeguards more precise so that legitimate requests are less likely to be blocked. In cybersecurity, the latest controls reportedly produce 60% fewer false positives than the previous system. The revised policy permits Fable 5.1 to help identify software vulnerabilities, although it does not allow the model to develop exploits for those weaknesses. For advanced biological capabilities, Anthropic has created a controlled-access program for Mythos 5.1 in partnership with the US government. Enrollment is expected to open to scientists, though the announcement does not provide a firm date.

The company supports its performance claims with results from coding, science, knowledge, and cost-efficiency evaluations, including Terminal-Bench-Science 0.1, Terminal-Bench 4.0, Humanity’s Last Exam, and CursorBench 3.2.0. Anthropic notes that the standard error on Terminal-Bench-Science is roughly 3.5 to 4.5 percentage points per model, meaning small score differences may reflect statistical noise. Its own evaluation setup reproduced public Claude Opus 5 and Fable 5 results within that uncertainty range. The cost charts compare models at multiple effort levels and use logarithmic scales, emphasizing the balance between capability and expenditure rather than presenting performance alone.

Because Fable 5.1 and Mythos 5.1 share the same underlying intelligence, Anthropic attributes observed benchmark differences between them to interventions by earlier, less precise cybersecurity safeguards. The company expects the gap to narrow after the newly announced safety improvements. This distinction matters because it suggests that restrictions, rather than core model quality, can materially affect measured performance on tasks that resemble security work.

Anthropic argues that Fable 5.1 is better at sustained reasoning and at resolving the underlying causes of software failures instead of taking shortcuts that merely hide symptoms. Its most striking example comes from investment firm Millennium. According to the account, an internal program suffered an extremely rare crash—about once per million executions—that engineers and other models had failed to explain over four to five years. Fable 5.1 reportedly disassembled a third-party vendor library, compared it with a core dump, and traced the failure to a defect inside that external library. The example is presented as evidence that the model can perform technically demanding diagnostic work whose manual cost would be difficult to justify.

Early-access partners cited both quantitative and qualitative improvements. One reported that Fable 5.1 solved more internal coding problems than Fable 5 or Opus 5, reached state-of-the-art performance on a trading-intuition evaluation, and remained readable during lengthy, multi-step assignments where earlier models became difficult to follow. Cognition said it planned to move Opus 5 traffic in its Devin coding agent to Fable 5.1 at launch because the new model matched or slightly exceeded Fable 5 in testing while reducing cost per task. Lower cache-read pricing reportedly made a Fable-class model economical for workloads previously assigned to Opus, beginning with code review.

Taken together, the announcement positions Fable 5.1 as a practical upgrade for agentic coding and complex professional work, while Mythos 5.1 serves as a restricted route to the same model’s most sensitive scientific and security capabilities. The claims remain based primarily on Anthropic’s benchmarks and partner testimony, but the combination of lower cached-context costs, customer-controlled safeguard infrastructure, fewer false positives, and stronger long-horizon problem solving could materially affect how organizations choose models for coding agents, research assistance, and enterprise automation.