SciCode-Verified: leaderboard
SciCode-Verified corrected v2 fresh-generation runs on 64 main problems and 287 subproblems, with background information, pass@1 and two-environment OR grading. The four-model run4 ledger is distinct from fixed-output regrading and from the historical public SciCode leaderboard. The paper contains matched controls; these separate leaderboard scores alone do not isolate the effect of benchmark repairs.
Metric: Subproblem Resolve Rate (%, corrected v2, with background, pass@1, two-environment OR). Source: github.com. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash | 97.2 | #66 |
| 2 | GPT-5.5 | 95.1 | #33 |
| 3 | Claude Opus 4.8 | 92.3 | #39 |
| 4 | GLM-5.2 | 84.3 | #80 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=scicode-verified · How It Works · Data refreshed daily, snapshot 2026-11-05.