RuBench (23-Task Freshness Gate): leaderboard
RuBench (2026) configurations whose model training cutoff leaves 23 of the 25 Russian-specified fix tasks eligible under the per-configuration freshness gate, scored on those 23 tasks. Pass@1 over three runs in percent, higher is better.
Metric: Pass@1 (%; the 23 tasks whose fix commits postdate the model training cutoff of these configurations; raw rows as ranked by the paper; pass@1, mean share of tasks resolved over three independent runs, judged by the withheld upstream maintainer regression tests; 25 repository-level fix tasks from aiohttp, aiogram, Laravel, NestJS and Fastify with task statements natively written in Russian; deployed agent CLI, model and reasoning effort as one configuration, product defaults otherwise, network access allowed). Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (xHigh) | 82.6 | #30 (GPT-5.6 Sol) |
| 2 | GPT-5.6 Luna (xHigh) | 75.4 | #121 (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=rubench-23-task-freshness-gate · How It Works · Data refreshed daily, snapshot 2026-11-13.