SEAL - SWE-Bench Pro (Private Dataset): leaderboard
Resolving bugs and implementing features in 276 private startup codebases never exposed publicly, scored on fix correctness without regressions. Measures generalization on unseen industrial code.
Metric: Score. Source: scale.com. Saturation forecast: Around January 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Muse Spark 1.1 | 51.5 | #52 |
| 2 | Claude Opus 4.6 (Thinking) | 47.1 | #63 (Claude Opus 4.6) |
| 3 | Muse Spark | 44.7 | #64 |
| 4 | GPT-5.4 (xHigh) | 43.4 | #79 (GPT-5.4) |
| 5 | Gemini 3.1 Pro (Preview) (Thinking) | 32.2 | #58 (Gemini 3.1 Pro (Preview)) |
| 6 | GPT-5.2 Codex | 27.74 | #95 |
| 7 | GPT-5.2 | 23.81 | #110 |
| 8 | Claude Opus 4.5 | 23.44 | #83 |
| 9 | Gemini 3 Pro | 17.95 | #80 |
| 10 | Claude Opus 4.1 | 17.75 | #140 |
| 11 | GPT-5 | 14.86 | #98 |
| 12 | Gemini 2.5 Pro (Preview) | 10.14 | #130 |
| 13 | Claude Sonnet 4 | 9.06 | #199 |
| 14 | GPT-4o | 3.62 | #343 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=seal-swe-bench-pro-private-dataset · How It Works · Data refreshed daily, snapshot 2026-10-20.