Summary
- On September 1, Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1, achieving scores of 52.6% on Terminal-Bench-Science 0.1 compared to Fable 5's 24.7%, and 55.8% on Terminal-Bench 4.0 against 42.0%.
- The cost of cache reads has been reduced by 75%, resulting in an overall workload cost decrease of about 25% and a reduction of up to 45% for highly agentic workloads, while basic pricing remains at $10/$50 per million tokens.
- Fable 5.1 is not part of Pro plans or standard Team seats, which operate on usage credits; however, Max and premium Team or Enterprise seats include it for up to 50% of weekly limits.
On Tuesday, Anthropic launched Claude Fable 5.1 alongside Claude Mythos 5.1, marking the first update to its Mythos-class models since the debut of Fable 5 on June 9.
The firm asserts that these new AI models are the world’s “most advanced for coding and knowledge work,” arriving three months into a launch cycle that has already experienced an 18-day export-control hiatus, a summer pricing dispute regarding subscription access, and the July rollout of Claude Opus 5, which underpriced Fable 5.
Myriad: Predict the release of GPT-6 from OpenAI. Make your prediction here.According to Anthropic, Fable 5.1 and Mythos 5.1 are based on the same model but utilize different safety filters. Fable 5.1 is accessible to all Claude account holders, while Mythos 5.1 is limited to approved cybersecurity and life sciences professionals through Anthropic's Cyber Verification Program and Life Sciences Verification Program, which succeeds the previous Project Glasswing access track for Mythos 5.
Understanding the Benchmarks
The key performance metric for Anthropic is derived from Terminal-Bench-Science 0.1, which evaluates an AI agent's capability to perform scientific research tasks in a command-line environment, measured as a pass rate.
Fable 5.1 achieved a score of 52.6%, significantly higher than Fable 5's 24.7% and Opus 5's 29.0%, effectively more than doubling its predecessor's performance.
In the Terminal-Bench 4.0 test, which assesses coding proficiency through terminal interactions—writing, executing, and debugging code across multiple command-line sessions—Fable 5.1 scored 55.8%, an increase from Fable 5's 42.0% and surpassing Opus 5's 52.3%.
Mythos 5.1, operating with less stringent cybersecurity filters, achieved a score of 60.9% on the same benchmark; Anthropic notes that the discrepancy is due to tasks that its safeguards redirected to Opus 4.8.
For the multidisciplinary reasoning test known as Humanity's Last Exam, which includes expert-level questions from various academic disciplines, Fable 5.1 scored 60.9% without external aids and 65.0% with them, outperforming Opus 5 and confirming its status as Anthropic's top model for academic applications.
Performance Comparison with Opus 5
Launched in July, Opus 5 offered a per-token price half that of Fable 5, while performing better than Fable 5 on most significant benchmarks, effectively making Fable 5 less relevant for many users.
Fable 5.1 has reversed this trend, now outperforming Opus 5 on all benchmarks published by Anthropic, including those where Opus 5 had previously exceeded Fable 5.
However, Fable 5.1 remains priced at twice the cost per token compared to Opus 5—$10 for input and $50 for output, versus Opus 5's $5 and $25. Tokens represent the basic units of information that a model processes (approximately two-thirds of an average English word), and companies incur costs based on usage, as intensive workflows can quickly exhaust subscription limits.
Anthropic promotes the model based on operational effort rather than merely raw scores, claiming that using Fable 5.1 at low or medium reasoning levels matches or exceeds Fable 5's previous results at a significantly lower cost, while reserving higher effort levels for particularly challenging problems.
For routine tasks, the rationale for opting out of Opus 5 diminishes as the effort level decreases.
Access to Fable 5.1 mirrors the structure established for Fable 5 after months of adjustments by Anthropic. For Pro plans and standard Team or Enterprise seats, Fable 5.1 draws exclusively from pay-as-you-go usage credits charged at the API rate since it is not included in the weekly limits of those plans. In contrast, Max plans and premium Team or Enterprise seats include it as a standard offering for up to 50% of weekly usage.
Claude Fable 5.1 is now available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry, identified by the model ID "claude-fable-5-1."