CursorBench 4.0 — Results
| Rank | Model | Score |
|---|---|---|
| 1 | claude-opus-5-5 | 57.8 |
| 2 | claude-sonnet-5-5 | 55.5 |
| 3 | claude-fable-5-1 | 51.8 |
| 4 | claude-opus-5 | 46.6 |
| 5 | grok-4-7 | 46.3 |
| 6 | gpt-5-6-sol | 41.7 |
| 7 | claude-sonnet-5 | 34.1 |
CursorBench 4.0
1 phaseActive
Cursor's updated internal agentic coding benchmark testing real Cursor coding-session tasks — a harder revision of the original CursorBench.
Overview
CursorBench 4.0
Quick answer: CursorBench 4.0 is Cursor's (Anysphere's) updated internal agentic coding benchmark, drawn from real Cursor coding-session tasks and evaluated inside Cursor's production agent harness. It is a harder, refreshed revision of the original CursorBench task set. Claude Opus 5.5 leads at 57.8% as of September 2026, with Claude Sonnet 5.5 close behind at 55.5%.
At a Glance
What it tests: A model's ability to complete real developer tasks sampled from live Cursor coding sessions, scored inside Cursor's own agent harness and tool-calling loop.
Why it matters: Because the task set is refreshed from live usage and the harness mirrors production Cursor, CursorBench 4.0 scores correlate closely with real-world in-IDE coding-agent quality rather than an isolated academic task format.
Known limitations: CursorBench 4.0 is proprietary to Cursor — the task set, harness, and scoring rubric are not publicly published, so scores are only directly comparable across models Cursor (or labs citing Cursor's harness) have evaluated.
What CursorBench 4.0 Measures
CursorBench 4.0 is the updated version of Cursor's (Anysphere's) internal benchmark for evaluating how well large language models perform as the reasoning engine behind its AI coding agent. Like its predecessor, it measures end-to-end task completion inside Cursor's own agent harness — multi-file edits, running tests, and iterating on tool output — but draws from a refreshed, harder set of tasks sampled from real Cursor coding sessions.
Because labs increasingly cite CursorBench 4.0 scores alongside their own model announcements (Anthropic and xAI both report it for their current-generation models), it has become a widely referenced proxy for real-world coding-agent usability, even though the underlying task set is not public and is distinct from the original, lower-saturation CursorBench.
Benchmark Specifications
| Field | Value |
|---|---|
| Task category | Coding agent |
| Metric | % score |
| Saturation | Low |
| Created by | Cursor (Anysphere) |
| Dataset | Proprietary — not publicly released |
| Project | cursor.com |
How CursorBench 4.0 Is Scored
Models are evaluated inside Cursor's agent harness on a refreshed suite of real developer tasks sampled from live coding sessions, and scored as a percentage reflecting successful task completion, verified by Cursor's internal grading pipeline.
State-of-the-Art Results
| Rank | Model | Score | Source | Date |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 57.8% | Anthropic: Claude Opus 5.5 | 2026-09 |
| 2 | Claude Sonnet 5.5 | 55.5% | Anthropic: Claude Sonnet 5.5 | 2026-09 |
| 3 | Claude Fable 5.1 | 51.8% | Anthropic: Claude Opus 5.5 | 2026-09 |
| 4 | Grok 4.7 | 46.3% (xhigh) | xAI: Grok 4.7 | 2026-09 |
| 5 | Claude Opus 5 | 46.6% | Anthropic: Claude Opus 5.5 | 2026-09 |
| 6 | GPT-5.6 Sol | 41.7% | Anthropic: Claude Opus 5.5 | 2026-09 |
| 7 | Claude Sonnet 5 | 34.1% | Anthropic: Claude Sonnet 5.5 | 2026-09 |
Scores sourced from Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5 announcements (September 2026) and xAI's Grok 4.7 announcement (September 2026). GPT-6 Astra and GPT-6 Sol scores were not publicly disclosed on this revision.
CursorBench 4.0 on Benchgen
No Benchgen results yet — be the first to run CursorBench 4.0.
CursorBench 4.0 vs Other Benchmarks
| Benchmark | What it tests | Saturation |
|---|---|---|
| CursorBench 4.0 | In-IDE agentic coding task completion (refreshed task set) | Low |
| CursorBench | In-IDE agentic coding task completion (original task set) | Low |
| DeepSWE | Real-world software engineering agent tasks | Low |
| Terminal-Bench 4.0 | Multi-step command-line agent tasks | Low |
Run CursorBench 4.0 on Your Model
Benchgen lets you track agentic coding performance across model and harness versions, complementing vendor-reported CursorBench 4.0 scores with independently repeatable evaluation.
Related Models & Benchmarks
- Claude Opus 5.5 — current CursorBench 4.0 leader
- Claude Sonnet 5.5
- Grok 4.7
- CursorBench — the original, lower-difficulty revision
- Terminal-Bench 4.0 — another agentic coding benchmark these models are compared on