Anthropic Launches Claude Fable 5.1 and Mythos 5.1, Cuts Cache Prices, Blocks Reasoning Copying

On September 1, 2026, Anthropic released two new models: the consumer-facing Claude Fable 5.1 and the institutional Claude Mythos 5.1. The company cut cache read pricing by three-quarters, reduced false safety blocks, and closed a loophole competitors had used to copy the model's reasoning process.

Fable 5.1 is now generally available under the name claude-fable-5-1 across the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.1, a hardened version built for institutions, remains restricted to verified organizations enrolled in the Project Glasswing program. Both models share the same underlying architecture but differ in the level and nature of the safeguards applied to them.

Cheaper Cache for Businesses

The most tangible change for business customers is cache read pricing, which fell from $1 to $0.25 per million tokens. Anthropic estimates typical office workloads will get about 25 percent cheaper as a result, while hours-long coding agent sessions that lean heavily on caching could cost up to 45 percent less. Base input and output token pricing stays the same at $10 and $50 per million tokens respectively, with lower batch-mode rates of $5 and $25.

The model keeps its million-token context window and a maximum output of around 128,000 tokens, with adaptive reasoning enabled by default on every query. Anthropic also introduced breaking API changes: forcing a specific tool via the tool_choice: any or tool parameters now returns a 400 error, intended to cut down on agents getting stuck making unnecessary calls.

No More Copying the Reasoning Trace

Anthropic closed a mechanism competitors had used to mass-query the API and train cheaper models on Claude's reasoning, a practice known as distillation. Until now, developers could manually edit earlier turns of a conversation while preserving the model's generated reasoning trace, letting them extract Fable 5's internal analytical process without paying the full compute cost. As of August 31, 2026, new API accounts have lost that ability: changing system prompts or tool definitions mid-conversation now invalidates the reasoning trace and forces it to be regenerated.

Anthropic also stipulated that reasoning traces are now tied to a specific model: Fable 5.1 can read the reasoning of earlier versions, but those earlier models cannot access traces generated by the newer version. It's an added barrier to reconstructing the model's reasoning chain outside Anthropic's own ecosystem.

Fewer False Blocks

The second practical problem Anthropic set out to fix was the previous version's overly cautious safety mechanisms, which often blocked legitimate requests. In the new version, cybersecurity safeguard interventions in Claude Code dropped by about 60 percent: the model still refuses to help build exploits, but is more permissive with vulnerability-detection tasks. For biology-related queries, such as medical or research questions, false blocks fell by 85 percent.

The model also introduces mandatory digital watermarking and C2PA metadata confirming the provenance of generated content, which Anthropic describes as the company's first model adaptation to the transparency requirements of the EU's AI Act.

Benchmark Results

Fable 5.1 more than doubled its predecessor's score on Terminal-Bench-Science 0.1, hitting 52.6 percent versus Fable 5's 24.7 percent; for comparison, Opus 5 scored 29.0 percent on the same test and GPT-5.6 Sol scored 22.4 percent. On Terminal-Bench 4.0 the model posted 55.8 percent, on CursorBench 3.2.0 it hit 73.4 percent, and on Humanity's Last Exam it scored 60.9 percent without tool use and 65.0 percent with it. The institutional Mythos 5.1 scored 60.9 percent on Terminal-Bench 4.0.

Anthropic also cites practical applications of the new model generation: model-generated GPU kernels for bioinformatics computations sped up seven tested models by 2.5 times, and mapping the surface of Venus from radar data reached a resolution of 2-3 kilometers, up from the previous 10-20 kilometers.

For Polish companies using Claude through Amazon, Google, or Microsoft cloud offerings, the cache price cut will be the most noticeable change, especially in agentic use cases where long working sessions generate large volumes of cache reads. The API changes requiring reasoning traces to regenerate after context edits, however, may require adjustments to existing integrations, particularly for companies building their own Claude-based developer tools.