Coding and agents
SWE-bench Verified
1 model with a published score, best first. Higher is better. The record has been broken 8 times since Oct 22, 2024.
96.0%Claude Opus 5Rank 1, published Jul 24, 2026
Leaderboard
Each model's latest published score. Every score links to its source.
| # | Model | SWE-bench V | Date | Note | Source |
|---|---|---|---|---|---|
| 1 | A Claude Opus 5RecordAnthropic | 96.0% | Jul 24, 2026 | lab-reported | benchlm.ai |
Record history
The best score ever published on SWE-bench Verified, and who held it. Each step is a new record.
| Date | Record holder | SWE-bench V | Gain | Note |
|---|---|---|---|---|
| Jul 24, 2026 | Claude Opus 5Current | 96.0% | +2.1 | lab-reported |
| Apr 7, 2026 | Claude Mythos Preview | 93.9% | +13.0 | lab-reported, gated preview |
| Nov 24, 2025 | Claude Opus 4.5 | 80.9% | +3.7 | lab-reported |
| Sep 29, 2025 | Claude Sonnet 4.5 | 77.2% | +2.3 | lab-reported |
| Aug 7, 2025 | GPT-5 | 74.9% | +2.2 | lab-reported |
| May 22, 2025 | Claude Sonnet 4 | 72.7% | +3.6 | lab-reported |
| Apr 16, 2025 | o3 | 69.1% | +6.8 | lab-reported |
| Feb 24, 2025 | Claude 3.7 Sonnet | 62.3% | +13.3 | lab-reported |
| Oct 22, 2024 | Claude 3.5 Sonnet | 49.0% | First | lab-reported |
Method
Lab-reported. OpenAI stopped reporting it in February 2026. Higher is better. Board source: swebench.com.
We collect the scores the board and the labs publish and run none ourselves. Ranks use each model's latest reading. 9 readings stored for this board, the newest from Jul 24, 2026. The record history includes older models that no longer have a page here.