AccelerateBench v1.1: leaderboard

HPC and scientific-computing program optimization benchmark (arXiv:2611.20882v1, Table 1) evaluating 16 frontier coding agents (at High reasoning effort in Claude Code or Codex) across 32 tasks from six supercomputing competitions. Scored by Score@1, the mean fraction of baseline runtime saved (%) on correct submissions.

Metric: Score@1 (mean runtime savings %). Source: arxiv.org. Saturation forecast: Around June 2028. 16 models tracked. Updated 21 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Sonnet 5.5High67.4#24
2Claude Opus 5.5High59.4#9
3GPT-6.1 SolHigh49.2#15
4GPT-6 AstraHigh47.2#10
5Hy4 previewHigh45.2#74
6hy4High45#40
7Kimi K3High42.5#61
8DeepSeek V4.1 Flash (1017)High40.1#47
9GLM-5.3High39.7#72
10GLM-5.3 FlashHigh39.3#133
11DeepSeek V4 ProHigh37.9#196
12GPT-6.1 TerraHigh37.9#25
13DeepSeek V4 FlashHigh34.1#267
14GLM-5.2High34#150
15Grok 4.6High33.9#50

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

Interactive version: theaggregate.ai/benchmark?slug=acceleratebench-v1-1 · How It Works · Data refreshed daily, snapshot 2026-11-21.