Claude Fable 5.1 Cuts Cache Costs for Long-Running Agents, but Upgrading Involves Breaking Changes to Thinking Blocks and Tool Calls

Anthropic launched Claude Fable 5.1 on September 1, positioning it for long-running agentic software development, scientific research, and multi-stage knowledge work. The model offers a 1 million-token context window and up to 128,000 output tokens, with the API identifier `claude-fable-5-1`. Input and output pricing remains $10 and $50 per million tokens, respectively, while cache reads have fallen to $0.25—one-quarter the price for Fable 5. Based on actual traffic from August, Anthropic estimates that costs could decline by roughly 25% for typical workloads and by as much as 45% for agentic tasks that rely heavily on context and tools. These figures, however, are vendor measurements rather than independent cross-platform evaluations.

The new release is more than a drop-in model replacement. The API adds per-message `effort`, system messages that apply only to the current turn, and `display: "updates"` for sending readable progress updates between tool calls. The migration documentation also identifies three breaking differences: the previous mode for requiring a specific tool to be called may now return an error outright; older models cannot read Fable 5.1 thinking blocks; and editing an earlier conversational turn invalidates existing thinking blocks. Systems that preserve complete agent trajectories, hand off tasks between models, or replay conversation history particularly need explicit version boundaries.

Anthropic's internal evaluations show Fable 5.1 scoring 55.8% on Terminal-Bench 4.0, compared with 42.0% for Fable 5. On Terminal-Bench-Science 0.1, it scored 52.6%, although that benchmark has a standard error of ±3.5 to 4.5 percentage points. Some safety evaluations may also assign a score of zero when a classifier intervenes or route the request to another Claude model, so the tables should not be treated as pure comparisons of model capability.

Engineering teams should next measure success rates, cache-hit rates, and cost per task using their own toolchains, while separately recording responses produced directly by Fable, responses produced by fallback models, and classifier refusals. Fable 5.1 is also subject by default to additional data-retention and safety-review policies. Although AWS began offering it at the same time, regional availability, data-governance requirements, and eligibility for zero data retention must still be verified individually.