Benchmarks · coding
SWE Atlas-QnA
Agents explore real repositories with shell tools and answer deep codebase questions; rubric-graded Task Resolve Rate (all rubric items met). Best harness per model/effort.
22 models · 22 results · data as of 2026-10-29 · labs.scale.com ↗
Leaderboard
Opus 5 xhigh
All results
22 rows| # | Model | Effort | SWE Atlas-QnA | Harness |
|---|---|---|---|---|
| 1 | Opus 5 | xhigh | 63.2% ±5.0 | Claude Code |
| 2 | Fable 5.1 | xhigh | 60.0% ±4.8 | Claude Code |
| 3 | GPT-6 Astra | xhigh | 59.1% ±4.9 | Codex |
| 4 | Opus 4.8 | xhigh | 57.3% ±4.9 | Claude Code |
| 5 | GLM 5.2 | max | 48.1% ±5.1 | Mini-SWE-Agent |
| 6 | Gemini 3.8 Flash | max | 47.0% ±5.1 | Mini-SWE-Agent |
| 7 | GPT-5.6 Sol | xhigh | 46.0% ±5.0 | Codex |
| 8 | GPT-5.5 | xhigh | 45.4% ±5.1 | Codex |
| 9 | Muse Spark 1.1 | xhigh | 42.2% ±5.1 | Mini-SWE-Agent |
| 10 | GPT-5.4 | xhigh | 40.8% ±4.9 | Mini-SWE-Agent |
| 11 | Opus 4.7 | max | 40.3% ±5.1 | Claude Code |
| 12 | Fable 5 | xhigh | 39.0% ±5.0 | Claude Code |
| 13 | Opus 4.6 | max | 33.3% ±5.0 | Claude Code |
| 14 | GPT-5.3 | xhigh | 32.6% ±4.9 | Codex |
| 15 | Sonnet 4.6 | max | 31.2% ±5.0 | Claude Code |
| 16 | DeepSeek V4 Pro | max | 27.1% ±4.7 | Mini-SWE-Agent |
| 17 | Muse Spark | 24.2% ±4.6 | ||
| 18 | GLM 5 | high | 20.5% ±4.5 | Mini-SWE-Agent |
| 19 | Gemini 3.1 Pro | high | 13.5% ±3.9 | Mini-SWE-Agent |
| 20 | Kimi K2.5 | max | 13.1% ±4.1 | Mini-SWE-Agent |
| 21 | MiniMax M2.5 | high | 10.3% ±3.5 | Mini-SWE-Agent |
| 22 | Gemini 3 Flash | high | 8.2% ±3.3 | Mini-SWE-Agent |