EvoCode-Bench: leaderboard
Coding agents are increasingly used as iterative development partners, but most benchmarks still evaluate one specification followed by one final assessment. This leaves out a basic question: can an agent keep its own codebase working as requirements change? We introduce EvoCode-Bench, a benchmark of 26 stateful coding tasks and 227 evaluated rounds. Each task preserves the agent's workspace for 5
Metric: MT@4 (self-reported). Source: benchmarklist.com. Saturation forecast: Around July 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 54 | #49 |
| 2 | GPT-5.5 | 52.4 | #31 |
| 3 | Claude Opus 4.6 | 44 | #63 |
| 4 | GLM-5.1 | 36.2 | #97 |
| 5 | Kimi K2.6 | 31.9 | #104 |
| 6 | Qwen 3.6 Plus | 29.4 | #120 |
| 7 | MiMo-V2.5 | 17.3 | #196 |
| 8 | Qwen 3.5 397B A17B | 4.6 | #146 |
| 9 | MiniMax-M2.7 | 3.7 | #258 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=evocode-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.