SWE-2 Coding Benchmark
How SWE-2 ranks on the LLM Coding Leaderboard, shown next to the three models directly above and below it, with the top three models included for reference.
Last evaluated: 4 weeks ago (October 1, 2026), with Devin.
| # | Model | Total points (max 70) |
Avg cost per prompt |
Avg time per prompt |
Video | Tested with | Points per project | |||
|---|---|---|---|---|---|---|---|---|---|---|
|
Laravel Code Quality
(max 20) |
React-TS Code Quality
(max 20) |
Bug Finding Test
(max 10) |
Edge-case Testing (4 projects, max 20) |
|||||||
| 1 | Opus 5.5 (High) | 67.26 | $0.83 | 03:42 |
Claude Code |
18.68 | 19 | 9.58 | 20 | |
| 2 | GPT-6-Astra (High) | 66.35 | $0.96 | 04:34 |
Codex CLI |
18.6 | 18.17 | 9.58 | 20 | |
| 3 | GPT-6-Astra (Medium) | 66.03 | $0.69 | 03:08 |
Codex CLI |
18.2 | 17.83 | 10 | 20 | |
| ⋯ 16 places not shown ⋯ | ||||||||||
| 20 | Haiku 5.5 (Medium) | 58.2 | $0.03 | 02:37 |
Claude Code |
17.32 | 17 | 6.88 | 17 | |
| 21 | Gemini 4 Argon (High) | 57.09 | N/A | 09:41 |
Devin |
17.15 | 18.5 | 7.19 | 14.25 | |
| 22 | GPT-6.1-Terra (Xhigh) | 57.09 | $0.07 | 01:29 |
Codex CLI |
16.23 | 17.5 | 8.96 | 14.4 | |
| 23 | SWE-2 (Max) | 57.07 | N/A | 11:59 |
Devin |
17.5 | 17.83 | 7.29 | 14.45 | |
| 24 | Mistral Large 4 (High) | 56.86 | $0.34 | 14:33 |
OpenCode |
18.13 | 19.33 | 3.65 | 15.75 | |
| 25 | Grok 4.6 (High) | 56.71 | $0.77 | 09:19 |
Cursor |
17.4 | 18 | 6.56 | 14.75 | |
| 26 | GPT-6-Luna (Xhigh) | 56.4 | $0.03 | 10:08 |
Codex CLI |
17.02 | 17.17 | 6.46 | 15.75 | |
Tutorials about SWE-2
Video
· Sep 15, 2026
I Tested NEW SWE-2 Model in Devin on 24 Coding Prompts
How to read this
- Each score measures a model-and-harness configuration, not the model in isolation.
- Prices are calculated with API costs. Models tested on a subscription plan show N/A.
- Full methodology and scoring formulas are explained in this article.
See every tested model side by side on the full LLM Coding Leaderboard.
