CursorBench
Real coding-agent sessions from Cursor, graded on whether the change does what the developer asked.
Agent Index weight
4%
in the Agentic coding pillar
Models tested
22
from the benchmark’s leaderboard and reports
Results
| # | Model | Score | Normalised | Harness | Cost/task | Date |
|---|---|---|---|---|---|---|
| 1 | Gemini 4 Carbon Supported Google | 60.4% | source | $9.22 | — | |
| 2 | Claude Fable 5.2 Supported Anthropic | 60.0% | source | $32.79 | — | |
| 3 | Claude Opus 5.5 Supported Anthropic | 57.8% | source | $13.43 | — | |
| 4 | Gemini 4 Hydrogen Estimated Google | 57.8% | source | $5.86 | — | |
| 5 | Claude Sonnet 5.5 Estimated Anthropic | 55.5% | source | $7.05 | — | |
| 6 | Gemini 4 Argon Supported Google | 52.9% | source | $4.57 | — | |
| 7 | Claude Fable 5.1 Supported Anthropic | 51.8% | source | $17.28 | — | |
| 8 | Muse Wonder Estimated Meta | 50.4% | source | $5.25 | — | |
| 9 | Claude Haiku 5.5 Estimated Anthropic | 48.4% | source | $1.12 | — | |
| 10 | DeepSeek V4.1 Flash 1017 Estimated DeepSeek | 47.7% | source | $0.98 | — | |
| 11 | Claude Opus 5 Supported Anthropic | 46.6% | source | $11.95 | — | |
| 12 | Grok 4.7 Estimated xAI | 46.3% | source | $6.01 | — | |
| 13 | GLM-5.3 Supported Z.ai (Zhipu) | 42.6% | source | $5.05 | — | |
| 14 | GPT-5.6 Sol Supported OpenAI | 41.7% | source | $8.23 | — | |
| 15 | Muse Spark 1.3 Estimated Meta | 41.6% | source | $2.64 | — | |
| 16 | Grok 4.6 Supported xAI | 41.4% | source | $6.10 | — | |
| 17 | GPT-5.6 Terra Estimated OpenAI | 41.3% | source | $5.14 | — | |
| 18 | Gemini 3.8 Flash Supported Google | 39.6% | source | $4.70 | — | |
| 19 | GLM-5.3-Flash Supported Z.ai (Zhipu) | 36.8% | source | $0.39 | — | |
| 20 | GPT-5.6 Luna Estimated OpenAI | 35.9% | source | $1.03 | — | |
| 21 | Claude Sonnet 5 Supported Anthropic | 34.1% | source | $7.17 | — | |
| 22 | Composer 2.5 Cursor | 27.7% | source | $0.68 | — |
For models without a CursorBench result, the index estimates one from Epoch’s Capabilities Index using a logistic fit on 20 models (residual SD 0.16 on the logit scale). Estimates are labelled and widen the model’s range.