CursorBench
Cursor's first-party benchmark for ambiguous, multi-file coding-agent tasks drawn from real Cursor sessions, currently on the 4.0 task set.
Top models on CursorBench — November 19, 2026
As of November 19, 2026, Gemini 4 Carbon leads the CursorBench leaderboard with 60.4% , followed by Claude Fable 5.2 (60.0%) and Claude Opus 5.5 (57.8%).
Gemini 4 Carbon
Claude Fable 5.2
Anthropic
Run by CursorBenchClaude Opus 5.5
Anthropic
Run by CursorBench21 modelsCoding10% of Coding reference weightCurrentUpdated November 19, 2026
Leaderboard (21 models)
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Gemini 4 CarbonGoogleRun by CursorBench | 60.4% | Run by CursorBench | Not reported | Closed |
| 2 | Claude Fable 5.2AnthropicRun by CursorBench | 60.0% | Run by CursorBench | Not reported | Closed |
| 3 | Claude Opus 5.5AnthropicRun by CursorBench | 57.8% | Run by CursorBench | Not reported | Closed |
| 4 | Claude Sonnet 5.5AnthropicRun by CursorBench | 55.5% | Run by CursorBench | Not reported | Closed |
| 5 | Gemini 4 HydrogenGoogleRun by CursorBench | 54.3% | Run by CursorBench | Not reported | Closed |
| 6 | Gemini 4 ArgonGoogleRun by CursorBench | 52.9% | Run by CursorBench | Not reported | Closed |
| 7 | Claude Fable 5.1AnthropicRun by CursorBench | 51.8% | Run by CursorBench | Not reported | Closed |
| 8 | Muse WonderMetaRun by CursorBench | 50.4% | Run by CursorBench | Not reported | Closed |
| 9 | Claude Haiku 5.5AnthropicRun by CursorBench | 48.4% | Run by CursorBench | Not reported | Closed |
| 10 | DeepSeek V4.1 Flash 1017DeepSeekRun by CursorBench | 47.7% | Run by CursorBench | Not reported | Open |
| 11 | Claude Opus 5AnthropicRun by CursorBench | 46.6% | Run by CursorBench | Not reported | Closed |
| 12 | Grok 4.7xAIRun by CursorBench | 46.3% | Run by CursorBench | Not reported | Closed |
| 13 | GLM-5.3Z.AIRun by CursorBench | 42.6% | Run by CursorBench | Not reported | Open |
| 14 | GPT-5.6 SolOpenAIRun by CursorBench | 41.7% | Run by CursorBench | Not reported | Closed |
| 15 | Muse Spark 1.3MetaRun by CursorBench | 41.6% | Run by CursorBench | Not reported | Closed |
| 16 | Grok 4.6xAIRun by CursorBench | 41.4% | Run by CursorBench | Not reported | Closed |
| 17 | GPT-5.6 TerraOpenAIRun by CursorBench | 41.3% | Run by CursorBench | Not reported | Closed |
| 18 | GLM-5.3-FlashZ.AIRun by CursorBench | 36.8% | Run by CursorBench | Not reported | Open |
| 19 | GPT-5.6 LunaOpenAIRun by CursorBench | 35.9% | Run by CursorBench | Not reported | Closed |
| 20 | Claude Sonnet 5AnthropicRun by CursorBench | 34.1% | Run by CursorBench | Not reported | Closed |
| 21 | Composer 2.5CursorRun by CursorBench | 27.7% | Run by CursorBench | Not reported | Closed |
- Lab-reported
- Benchmark-run
- Independent
According to BenchLM.ai, Gemini 4 Carbon leads the CursorBench benchmark with a score of 60.4%, followed by Claude Fable 5.2 (60.0%) and Claude Opus 5.5 (57.8%). The top models are clustered within 2.6 points, suggesting this benchmark is nearing saturation for frontier models.
21 models have been evaluated on CursorBench. The benchmark falls in the Coding category. BenchAlign v5.8 gives CursorBench 10% of the Coding reference weight, so it moves the Coding leaderboard and the overall ranking. Reference weights are relative weights in the calibrated model, not fixed shares of a score.
About CursorBench
Year
2026
Tasks
Long-horizon multi-file agentic coding tasks
Format
Cursor agent-loop evaluation
Difficulty
Professional agentic software engineering
Cursor published the CursorBench 4.0 task set on September 10, 2026, adding long-horizon edit, refactor, investigation, intent-understanding, job-management, and design-adherence problems; 4.0 scores are not comparable with the retired 3.2 set. BenchLM maps the highest published effort tier for each model. Since November 12, 2026, 4.0 is the CursorBench instrument in the coding score; the frozen CursorBench 3.2 rows stay on model pages for reference.
Freshness and provenance
Version
CursorBench 4.0
Refresh cadence
Continuous
Staleness state
Current
Question availability
Tasks private; drawn from real Cursor sessions
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does CursorBench measure?
Cursor's first-party benchmark for ambiguous, multi-file coding-agent tasks drawn from real Cursor sessions, currently on the 4.0 task set.
Which model scores highest on CursorBench?
Gemini 4 Carbon by Google currently leads with a score of 60.4% on CursorBench.
How many models are evaluated on CursorBench?
21 AI models have published results on CursorBench in the BenchLM catalog.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.