All tests

Agents

CursorBench

Ambiguous, multi-file coding tasks drawn from real sessions in the Cursor editor, run and scored by Cursor.

Published by Cursor CursorBench, compiled by Epoch AI under CC BY 4.0. Compare it with other tests

RSS feed
Top score57.8%Claude Opus 5.5 (max), Anthropic
Models tested14from 6 makers
Median score41.6%Half the models score above it
Scale0 to 100%The share of tasks done right

Top 14

Each model at its best configuration, coloured by maker.

The best model from each maker

Each maker’s highest-scoring model on this test, and its rank.

MakerBest modelScoreRank
AnthropicClaude Opus 5.5 (max)57.8%#1 of 14
xAIGrok 4.7 (xhigh)46.3%#5 of 14
Z.aiGLM-5.3 (max)42.6%#6 of 14
OpenAIGPT-5.6 Sol (max)41.7%#7 of 14
Meta AIMuse Spark 1.3 (max)41.6%#8 of 14
GoogleGemini 3.8 Flash (high)39.6%#11 of 14

New top scores

Each time a different model took the top spot, since the site began keeping this record.

  1. Claude Opus 5.5 took the top spot at 57.8% from Claude Fable 5.1, which had led at 51.8%.

Every result

14 models, each once at its best configuration, highest score first. Models with the same score share a rank.

RankModelMakerScore
#1Claude Opus 5.5 (max)Anthropic57.8%
#2Claude Sonnet 5.5 (max)Anthropic55.5%
#3Claude Fable 5.1 (max)Anthropic51.8%
#4Claude Opus 5 (max)Anthropic46.6%
#5Grok 4.7 (xhigh)xAI46.3%
#6GLM-5.3 (max)Z.ai42.6%
#7GPT-5.6 Sol (max)OpenAI41.7%
#8Muse Spark 1.3 (max)Meta AI41.6%
#9Grok 4.6 (xhigh)xAI41.4%
#10GPT-5.6 Terra (max)OpenAI41.3%
#11Gemini 3.8 Flash (high)Google39.6%
#12GLM-5.3-Flash (max)Z.ai36.8%
#13GPT-5.6 Luna (max)OpenAI35.9%
#14Claude Sonnet 5 (max)Anthropic34.1%

How it is scored

Scores, average cost, tokens and steps per task as published by Cursor and compiled by Epoch AI. Cost applies list prices per million tokens to each model’s token use. A suffix such as _high names the reasoning effort.

Results from Cursor CursorBench, compiled by Epoch AI (CC BY 4.0), as published. Model names link to their pages where the site can match them; makers come from the model’s page, or else from the publisher.