"Grok 4.5 has an advantage on CursorBench: an earlier snapshot of the Cursor codebase was unintentionally included in training"

# Cursor · CursorBench Source: [https://cursor.com/cursorbench](https://cursor.com/cursorbench) We evaluate agents on ambiguous, multi\-file tasks from real Cursor sessions\. Higher scores are better\. [More about CursorBench](https://cursor.com/blog/cursorbench)A scatter and line chart comparing Fable 5, Opus 4\.8, GPT\-5\.5, Sonnet 5, Grok 4\.5, GLM 5\.2, Composer 2\.5, 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 taskGrok 4\.5 high[\*](https://cursor.com/cursorbench#cursorbench-grok-note)Fable 5 highOpus 4\.8 highSonnet 5 highComposer 2\.5GPT\-5\.5 mediumGLM 5\.2 highKimi K2\.7 CodeGemini 3\.5 FlashModel1Fable 5 Max70\.5%$17\.32103,525722Fable 5 Extra High68\.4%$11\.7364,971563Grok 4\.5 High[\*](https://cursor.com/cursorbench#cursorbench-grok-note)66\.7%$1\.5119,521334Fable 5 High66\.5%$8\.7743,747485Grok 4\.5 Medium[\*](https://cursor.com/cursorbench#cursorbench-grok-note)65\.4%$1\.5418,914346Fable 5 Medium65\.2%$6\.8030,366417Grok 4\.5 Low[\*](https://cursor.com/cursorbench#cursorbench-grok-note)63\.5%$1\.2215,841318Opus 4\.8 Max62\.3%$5\.7771,411449Fable 5 Low62\.1%$4\.4618,1823110Sonnet 5 Max61\.5%$6\.4592,8828611Opus 4\.8 Extra High59\.4%$4\.5051,1214012Sonnet 5 Extra High58\.7%$4\.1652,8716713GPT\-5\.5 High58\.4%$2\.0512,1832814GPT\-5\.5 Extra High58\.4%$2\.8517,5343215Opus 4\.8 High58\.0%$3\.1533,5483316Sonnet 5 High56\.9%$3\.1939,4835717Opus 4\.8 Medium56\.1%$2\.8128,3843218Composer 2\.556\.1%$0\.4414,2863319GLM 5\.2 Max55\.0%$1\.7635,9465820GPT\-5\.5 Medium53\.8%$1\.518,5222521Opus 4\.8 Low53\.1%$2\.0219,6242722Sonnet 5 Medium52\.4%$2\.1626,2004623GLM 5\.2 High51\.5%$1\.1921,8294924Kimi K2\.7 Code49\.7%$1\.4331,2475825Gemini 3\.5 Flash48\.8%$2\.2046,7027726Sonnet 5 Low47\.7%$1\.3016,2693327GPT\-5\.5 Low46\.6%$0\.985,16820 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](https://cursor.com/blog/grok-4-5)\. ## Changelog ### CursorBench 3\.2 - Introduced instruction following and advanced tool use problems\. ### CursorBench 3\.1 - Introduced problems focused on codebase understanding, bugfinding, planning, and code review\. - Improved grading criteria for some edit 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](https://cursor.com/docs/models-and-pricing)\(input, cache read, cache write, and output\) to the tokens it used on each task across the CursorBench 3\.2 benchmarks, then averaging with the same task weights as the CursorBench 3\.2 score\. Results are subject to variance; small differences in scores may not be statistically meaningful\.