Cursor Adds Claude Fable 5.1, Pairing a 73.4% Coding Score With Max-Effort Costs

Cursor has added Anthropic’s Claude Fable 5.1 to its model picker, bringing a new option built around prolonged coding work into the editor. The model recorded 73.4% on CursorBench 3.2 at maximum effort, ahead of Fable 5 and Opus 5 in that evaluation—but the same effort setting can demand far more time and tokens than lighter modes.

The release gives Cursor users multiple effort levels, a thinking variant and a 300,000-token context window in settings. Fable 5.1 is also generally available through AWS, extending the model beyond Anthropic’s own routes. Cursor users with Privacy Mode enabled must first approve Fable 5.1’s data-retention policy in the Cursor Dashboard.

A completion claim that depends on checking

Cursor presents self-verification as the important capability behind the score. The intended loop is not simply to write a patch: the agent should continue testing its assumptions and inspecting whether the work holds together before declaring it finished. That matters most when a task crosses files or runs long enough for a model to mistake partial progress for a completed job.

That is a product claim, not a demonstrated guarantee that every generated change will be sound. CursorBench 3.2 is the company’s coding evaluation, and its leading result is explicitly measured at maximum effort. The score establishes a controlled comparison with the named models, while leaving the economics and reliability of ordinary production use to users’ own workloads.

73.4%Claude Fable 5.1

Fable 5.1 scored 73.4% on CursorBench 3.2 at maximum effort.

70.5%Fable 5

Fable 5 scored 70.5% on the same CursorBench 3.2 comparison.

70.0%Opus 5

Opus 5 scored 70.0% on the same CursorBench 3.2 comparison.

Reasoning is the adjustable resource

Fable 5.1 offers five reasoning levels: low, medium, high, xhigh and max. There is no setting that turns reasoning off entirely. The design gives developers a way to trade latency and spending for more deliberate work, rather than treating the model’s strongest mode as the default for every request.

A hands-on SVG experiment by Simon Willison makes that tradeoff unusually visible. On the same request to generate a pelican riding a bicycle, low and medium used roughly 2,000 output tokens and showed no summarized reasoning text. High used 2,612 output tokens, took 29.6 seconds and cost about 13 cents.

At xhigh, that one prompt grew to 36,767 output tokens, ran for 7 minutes 51 seconds and cost $1.83. At max, it reached 65,927 output tokens over 13 minutes 54 seconds for $3.30. Willison judged the max result his best pelican from an Anthropic model, but the test is an illustration of reasoning behavior, not a measure of software-engineering success.

Where the model’s promise is strongest—and least settled

• The clearest published coding result is the 73.4% CursorBench score at maximum effort, not a score for the lower-cost settings.
• Anthropic’s broader launch figures also point to gains on agent-style tasks: it reported 52.6% on Terminal-Bench-Science 0.1, versus 24.7% for Fable 5.
• Browserbase reported an 82% completion rate for Fable 5.1 on its hardest browser-agent benchmark, versus 74% for Opus 5 and 57% for Fable 5. That is partner-supplied evidence from a different kind of agent task.

For Cursor users, then, Fable 5.1 is not merely another model name in a picker. It is a choice to spend more computation on a job whose route is uncertain: difficult debugging, broad changes across a codebase, or a task an agent must pursue without constant supervision. The immediate test is whether its checking loop avoids enough rework to offset the extra time and token bill.