CursorBench 3.2

We evaluate agents on ambiguous, multi-file tasks from real Cursor sessions. Higher scores are better.

More about CursorBench
A scatter and line chart comparing Fable 5, Opus 5, Opus 4.8, GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, GPT-5.5, Sonnet 5, Grok 4.5, GLM 5.2, Composer 2.5, Gemini 3.6 Flash, Gemini 3.5 Flash, and Kimi K2.7 Code scores against average cost per task.75%CursorBench 3.2 score70%65%60%55%50%45%$20$16$12$8$4$0Average cost per taskOpus 5 highGrok 4.5 high*Fable 5 highGPT-5.6 Sol mediumSonnet 5 highComposer 2.5Gemini 3.6Flash high
Model
1Fable 5 Max70.5%$17.32103,52572
2Opus 5 Max70.0%$8.2361,83878
3Opus 5 Extra High69.3%$7.3554,23972
4Fable 5 Extra High68.4%$11.7364,97156
5GPT-5.6 Sol Max67.2%$5.6928,32048
6Opus 5 High66.7%$3.9127,93248
7Grok 4.5 High*66.7%$1.5119,52133
8Fable 5 High66.5%$8.7743,74748
9Grok 4.5 Medium*65.4%$1.5418,91434
10Fable 5 Medium65.2%$6.8030,36641
11GPT-5.6 Terra Max64.9%$2.8932,96947
12GPT-5.6 Sol Extra High64.5%$3.8819,69938
13Opus 5 Medium64.3%$3.2923,61244
14Grok 4.5 Low*63.5%$1.2215,84131
15GPT-5.6 Sol High63.5%$2.7913,86732
16Opus 5 Low62.8%$2.5518,52937
17Opus 4.8 Max62.3%$5.7771,41144
18Fable 5 Low62.1%$4.4618,18231
19Sonnet 5 Max61.5%$6.4592,88286
20GPT-5.6 Luna Max61.1%$1.9787,97361
21GPT-5.6 Sol Medium60.0%$1.959,74727
22Opus 4.8 Extra High59.4%$4.5051,12140
23GPT-5.6 Terra Extra High59.2%$1.4416,08929
24Sonnet 5 Extra High58.7%$4.1652,87167
25GPT-5.5 High58.4%$2.0512,18328
26GPT-5.5 Extra High58.4%$2.8517,53432
27Opus 4.8 High58.0%$3.1533,54833
28GPT-5.6 Luna Extra High57.7%$1.1422,48048
29Sonnet 5 High56.9%$3.1939,48357
30GPT-5.6 Luna High56.8%$0.8215,14140
31Opus 4.8 Medium56.1%$2.8128,38432
32Composer 2.556.1%$0.4414,28633
33GLM 5.2 Max55.0%$1.7635,94658
34GPT-5.6 Terra High54.2%$0.899,46823
35GPT-5.5 Medium53.8%$1.518,52225
36Gemini 3.6 Flash High53.5%$1.5630,43664
37Opus 4.8 Low53.1%$2.0219,62427
38GPT-5.6 Sol Low52.6%$1.015,10419
39Sonnet 5 Medium52.4%$2.1626,20046
40GLM 5.2 High51.5%$1.1921,82949
41Gemini 3.6 Flash Medium51.2%$1.4828,51162
42GPT-5.6 Terra Medium50.3%$0.616,22220
43Kimi K2.7 Code49.7%$1.4331,24758
44Gemini 3.5 Flash48.8%$2.2046,70277
45GPT-5.6 Luna Medium47.7%$0.397,09528
46Sonnet 5 Low47.7%$1.3016,26933
47Gemini 3.6 Flash Low47.4%$1.1320,52950
48GPT-5.6 Terra Low46.9%$0.535,31219
49GPT-5.5 Low46.6%$0.985,16820
50GPT-5.6 Luna Low37.6%$0.163,20917

Grok 4.5 has an advantage on CursorBench: an earlier snapshot of the Cursor codebase was unintentionally included in training. The exact score impact is unclear. That data has been removed for future models. For a rundown of third-party benchmark scores, see the Grok 4.5 launch blog.

Changelog

Reporting

  • Updated GPT-5.6 Sol, Terra, and Luna results to account for cache write costs.

Tasks

  • CursorBench 3.2
    • Introduced instruction following and advanced tool use problems.

Tasks

  • CursorBench 3.1
    • Introduced problems focused on codebase understanding, bugfinding, planning, and code review.
    • Improved grading criteria for some edit tasks.

Tasks

  • CursorBench 3.0
    • Initial set of tasks focused on edit, refactor, and bugfix problems.

Avg cost / task is computed by applying each model's published per-million-token pricing (input, cache read, cache write, and output) to the tokens it used on each task. Results are subject to variance; small differences in scores may not be statistically meaningful.