MERA Code - JavaTestGen: leaderboard
MERA Code (Russian) JavaTestGen task: generating Java unit tests for supplied code. Each task keeps its own metric — pass rates, CodeBLEU and chrF all appear across this suite — so read the metric label before comparing boards.
Metric: pass@1 (%). Source: mera.a-ai.ru. Saturation forecast: Around April 2027. 66 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 62.56 | #79 |
| 2 | GPT-5.1 Codex Max | 55.95 | #100 |
| 3 | Kimi K2.5 | 37.44 | #144 |
| 4 | GLM-5.1 | 37.44 | #97 |
| 5 | DeepSeek V4 Pro | 36.12 | #102 |
| 6 | Qwen 3.5 27B | 35.68 | #187 |
| 7 | Claude Opus 4.6 | 34.8 | #63 |
| 8 | GPT-4.1 | 34.36 | #248 |
| 9 | Qwen 3.6 27B | 34.36 | #170 |
| 10 | DeepSeek V4 Flash | 32.16 | #137 |
| 11 | Qwen 3 30B A3B 2507 (Thinking) | 31.72 | #376 (Qwen 3 30B A3B 2507) |
| 12 | Qwen 3.5 122B A10B | 31.28 | #173 |
| 13 | Qwen 3.6 35B A3B | 30.84 | #226 |
| 14 | GPT-OSS-120B | 27.31 | #341 |
| 15 | Qwen 3 32B | 25.99 | #443 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mera-code-javatestgen · How It Works · Data refreshed daily, snapshot 2026-10-20.