Benchmarks
357 sources · 1,263 models × 3,079 benchmarks · 162,127 scoresCursorBench 4.1 evaluates agents on ambiguous, multi-file tasks from real Cursor sessions. Released November 18, 2026, this version modifies CursorBench 4.0's instruction-following problems and increases the agent time limit, and every run carried over from 4.0 was rescored (by -4.9 to +3.6 points), so it is kept apart from 4.0. The public table reports task scores in percentage points without disclosing an exact evaluation item count.
84th strongest of 818 fitted benchmarks
Signal-to-noise +14.40
194th hardest of 818 fitted benchmarks
Capability needed 165.4
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Gemini 4 Neon | x-high | 65.4 | |
| 2 | Gemini 4 Carbon | high | 61.5 | |
| 3 | Gemini 4 Hydrogen | x-high | 61.4 | |
| 4 | Claude Fable 5.2 | max · Cursor production agent harness, with the prompt and harness improvements Cursor credits for part of Fable 5.2's score | 61.3 | |
| 5 | Claude Opus 5.5 | max | 58.1 | |
| 6 | Gemini 4 Argon | high | 56.5 | |
| 7 | Claude Sonnet 5.5 | max | 54.4 | |
| 8 | Grok 4.7 | x-high | 48.6 | |
| 9 | Claude Haiku 5.5 | max | 48.4 | |
| 10 | Grok 4.6 | x-high | 44.3 | |
| 11 | Muse Spark 1.3 | max | 41.9 | |
| 12 | GPT-5.6 Sol | max | 40.9 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).