Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, pairing benchmark gains in coding and reasoning with a 75% reduction in cache-read costs and the company's first text-output watermarking system1,2,4,9.
Both models are priced at $10 input / $50 output per million tokens, matching Claude Fable 5's pricing, but prompt cache reads drop to $0.25 per million tokens, or 0.025× the base input price, compared with 0.1× on prior models11. Each model supports a one-million-token context window, 128,000 max output tokens, and always-on adaptive thinking.
Claude Fable 5.1 is available through the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Claude Mythos 5.1 is restricted to registered Anthropic partners in cybersecurity or life sciences research, distributed through what Anthropic calls Project Glasswing.
On benchmarks, Anthropic said the models set records on Terminal-Bench 4.0 for CLI-based coding and Humanity's Last Exam for general reasoning. Claude Fable 5.1 scored 52.6% on Terminal-Bench-Science6. Anthropic described programming benchmark performance as doubling over the prior generation. The company said Fable 5.1 sets new standards on coding, knowledge work, and long-running problem-solving tasks, and can fix root causes of software issues10.
The Fable 5.1 release also includes changes to reduce false-positive restrictions from the model's safeguards. Anthropic released three scientific findings generated by the models before launch, including a custom GPU optimization and a high-resolution map of Venus assembled from existing photographs.
Claude Fable 5.1 and Claude Mythos 5.1 are Anthropic's first models to watermark text outputs, adding invisible watermarks to all text and file outputs. A detection API is available to eligible groups as required under EU law. Supported image and video files produced through the code execution tool carry C2PA Content Credentials when retrieved through the Files API.
Anthropic's system card for Mythos 5.1 rates the model as low-risk for concerns related to automated AI development, stating that its ability to accelerate internal AI R&D progress is in line with current trends. The card notes Mythos 5.1 is slightly more prone to general misbehavior than Opus and cooperates with human misuse somewhat more readily than Opus 5, but is less likely than previous models to ignore explicit constraints, hallucinate inputs, or falsely claim to have completed tasks.
On data handling, Anthropic has embraced Zero Data Retention, allowing clients to run models on their own infrastructure without data outflows. A high-privacy service called Enterprise Frontier Safeguards will extend this to Fable, though the system will still monitor for misuse, with clients controlling how monitoring takes place. Anthropic stated that it has never trained on enterprise data without explicit permission.
API-level changes include the removal of tool_choice types "any" and "tool" on both new models, which now return a 400 error; "auto" and "none" remain unchanged. Thinking blocks produced by the new models are preserved only for the model that produced them or a newer one; earlier models cannot read them.
The launch arrived a day after Anthropic inked a $35 billion infrastructure deal with Lambda Inc. and roughly a week after signing an even larger hardware contract with Nscale Global Holdings Ltd.3[2].