SWE-Serve: leaderboard
SWE-Serve (2026) evaluates coding agents on production inference-serving engineering with 53 repository-level tasks derived from merged SGLang changes, each scored by hidden functional, regression, end-to-end serving and performance tests under a closed-book network policy. The metric is mean pass@1 over three complete benchmark runs, in percent, higher is better; every model is run under the mini-SWE-agent harness and shown at each reasoning-effort setting evaluated.
Metric: pass@1 (%; mean over three runs of the fraction of the 53 tasks whose agent patch passes every hidden functional, regression, E2E and performance test; mini-SWE-agent harness, closed-book). Source: arxiv.org. Saturation forecast: Around February 2027. 31 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 75.5 | #29 (GPT-5.6 Sol) |
| 2 | Claude Opus 5 (Max) | 75.5 | #20 (Claude Opus 5) |
| 3 | Claude Opus 5 (High) | 73.6 | #20 (Claude Opus 5) |
| 4 | Claude Opus 5 (xHigh) | 72.3 | #20 (Claude Opus 5) |
| 5 | Claude Opus 5 (Medium) | 68.6 | #20 (Claude Opus 5) |
| 6 | GPT-5.6 Sol (xHigh) | 66 | #29 (GPT-5.6 Sol) |
| 7 | Kimi K3 (Max) | 64.2 | #53 (Kimi K3) |
| 8 | GPT-5.6 Luna (Max) | 64.2 | #117 (GPT-5.6 Luna) |
| 9 | GPT-5.6 Terra (Max) | 64.2 | #60 (GPT-5.6 Terra) |
| 10 | Claude Sonnet 5 (xHigh) | 64.2 | #108 (Claude Sonnet 5) |
| 11 | Claude Sonnet 5 (Max) | 61 | #108 (Claude Sonnet 5) |
| 12 | Claude Opus 5 (Low) | 60.4 | #20 (Claude Opus 5) |
| 13 | Claude Sonnet 5 (High) | 57.9 | #108 (Claude Sonnet 5) |
| 14 | GPT-5.6 Sol (High) | 56 | #29 (GPT-5.6 Sol) |
| 15 | GPT-5.6 Luna (xHigh) | 55.3 | #117 (GPT-5.6 Luna) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=swe-serve · How It Works · Data refreshed daily, snapshot 2026-11-12.