EvoCode-Bench - Single-Round: leaderboard

EvoCode-Bench (2026) single-round protocol: before each of 227 rounds of 26 evolving coding tasks, the workspace is fast-forwarded to the reference solution, and the agent (each model in the Terminus-2 harness) must pass the cumulative tests of that round in one attempt. Pass rate in percent; higher is better.

Metric: Single-round pass rate (SR, %): share of the 227 rounds solved in one attempt after Harbor fast-forwards the workspace to the reference solution of all earlier rounds, over the 26 stateful multi-round coding tasks of EvoCode-Bench (227 rounds, 5 to 15 per task, requirements that extend, correct or contradict earlier ones, cumulative executable verifiers), each model driving the Terminus-2 agent harness in a Harbor Docker workspace; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.678.9#113
2Claude Opus 4.7High76.7#85
3GPT-5.5High74.4#54
4GLM-5.163.9#198
5Kimi K2.659#192
6Qwen 3.6 PlusThinking57.3#238
7DeepSeek V4 ProHigh56.4#196
8Gemini 3.1 Pro (Preview)High46.7#99
9DeepSeek V4 FlashHigh46.3#267
10Qwen 3.5 397B A17B44.1#264
11MiniMax-M2.730#539
12Seed 2.0 ProHigh23.8#222
13MiMo-V2.5-ProHigh7.9#284

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=evocode-bench-single-round · How It Works · Data refreshed daily, snapshot 2026-11-21.