CursorBench: Sonnet 4.6 Scores 49%, as a Baseline for Sonnet 5 Launch Comparison

One-sentence takeaway

Cursor officially reported Claude Sonnet 5 at 57% and Claude Sonnet 4.6 at 49% on CursorBench; this shows 4.6 remains the comparison baseline for Cursor's internal coding agent evaluation, but the public post did not provide questions, configuration, or per-question trajectories.

Test environment

• Task: CursorBench (Cursor's proprietary coding/agent evaluation; the post does not elaborate on the definition).

• Comparison models: Claude Sonnet 5 vs Claude Sonnet 4.6.

• Release context: Announcing that Claude Sonnet 5 is now available in Cursor.

• Visible metrics: 57% vs 49%.

Inputs/configuration

The X post did not disclose prompts, agent harness, effort, tool permissions, number of repetitions, or whether thinking was enabled. These two percentages cannot be treated as equivalent scores on SWE-bench or other public leaderboards.

Results data

• Claude Sonnet 5: 57% on CursorBench.

• Claude Sonnet 4.6: 49% on CursorBench.

• The author described the improvement as "significant".

• Post engagement visible: approximately 1.159 million views, 205 reposts, 349 replies, 4,179 likes (figures at collection time).

Conclusion

In Cursor's own agent evaluation, Sonnet 4.6 is clearly lower than the subsequent Sonnet 5. If your workflow is tied to Cursor, 4.6 is better suited as a "validated, cheaper baseline" rather than the latest strongest Sonnet on that client. If your workflow is not in Cursor, this result only indicates the relative gap and cannot alone determine whether to keep using 4.6.

Limitations

• Only two total scores, no breakdowns, variance, or failure types.

• The CursorBench task set was not disclosed in this post, so it cannot be independently reproduced.

• Vendor-provided comparisons when launching new products carry baseline-selection bias.

• Whether 4.6 and 5 used the same effort, context window, and tool set was not stated.

Reproduction steps

1. In Cursor, separately lock Claude Sonnet 4.6 and Sonnet 5, disabling automatic upgrades to other models.

2. Prepare the same set of real PRs/issues, saving complete agent trajectories, diffs, and test results.

3. Score based on "first-pass success, green tests, no manual rework" rather than simply copying 49/57.

4. Report Cursor version, mode (Ask/Agent), effort, and whether 1M context is enabled.