On September 1, 2026, Anthropic released Claude Fable 5.1 in the official Claude Platform version notes, positioning it as long-running agent programming, knowledge work, and research, with the model identified as claude-fable-5-1. On the same day, Claude Mythos 5.1 was only available to Project Glasswing participants and should not be confused with the publicly available Fable 5.1.
The upgrade focus is not just on the 1 million token context
Fable 5.1 provides a default 1 million token context, up to 128,000 token outputs, and continuously enabled adaptive thinking. It has been introduced to Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry, suitable for teams crossing large files, using tools for long periods, or maintaining task states.
Longer context does not mean you can stack data infinitely. Input sorting, tool return clipping, historical compression, and permission isolation still determine whether long tasks are stable. Existing projects can first use the site's Context Caching Cost Explanation to estimate duplicate inputs and then decide which content is worth long-term retention.
When migrating, first check three compatibility points
- tool_choice's any and tool are not supported on Fable 5.1 and Mythos 5.1, and requests return 400; existing code should be switched to auto or none, or strict tool calls and structured output must ensure parameters conform to the schema.
- Thinking block cannot arbitrarily replay across old models. For new accounts created on or after August 31, 2026, if system prompts, tool lists, or earlier messages were modified before playback, the interface may also show errors.
- Effort-based message adjustments, single-turn system messages, and progress updates are still in beta and require corresponding beta request headers. You cannot assume that all cloud channels, SDKs, and accounts are synchronously supported.
These changes are more likely to cause real failures than replacing model names. Claude Code or custom programming agents often encapsulate tool selection and historical replay within the framework; request logs and error handling should be checked before migration. When context suddenly expands, refer to Claude Code's context troubleshooting method.
Costs haven't increased, but cache read prices have dropped significantly
The input and output prices of Fable 5.1 remain at $10 and $50 per million tokens, the same as Fable 5; Prompt cache read drops to $0.25 per million tokens, with write prices unchanged. Stable prefixes and duplicate data are more worth caching, but only when the hit rate is high enough can real savings be achieved.
Another boundary is 30-day data retention: unless explicitly authorized by Anthropic, Fable 5.1 does not provide zero data retention. Teams should first conduct regression testing with representative long tasks, confirming that tool calls, cache hits, thinking block playbacks, and data requirements all pass, then gradually expand traffic.
