DataCenterBench Tier 1
Agents · 2026-10-28
Robocurve's benchmark of frontier models driving two real robot arms through ten tabletop steps rebuilt from data-center equipment manuals, such as flipping a switch or joining Ethernet connectors. Each model gets three tries per task, watching through cameras at 224 by 224 pixels, with at most 40 model calls per try. Two LLM judges score each try from photos on 0/1 or 0/½/1 rubrics, and a person settles disagreements. avg@3 is the mean of three runs of the ten tasks. Robocurve says models that think longer overheat the motors more, and scores a few points apart aren't a reliable ranking.
Benchmark source: Robocurve
Top models (higher is better)
| Model | Score | Source |
|---|---|---|
| GPT-6 Astra | 40.0 | DataCenterBench Tier 1: GPT-6 Astra · ultrafast · Robocurve |
| GPT-6.1 Sol | 38.3 | DataCenterBench Tier 1: GPT-6.1 Sol · ultrafast · Robocurve |
| Gemini 4 Argon | 35.0 | DataCenterBench Tier 1: Gemini 4 Argon · Robocurve |
| Claude Opus 5.5 | 21.7 | DataCenterBench Tier 1: Claude Opus 5.5 · Robocurve |
| Muse Wonder | 16.7 | DataCenterBench Tier 1: Muse Wonder · Robocurve |
| GLM-5.3 Flash | 15.0 | DataCenterBench Tier 1: GLM 5.3 Flash · Robocurve |
| Kimi K3 | 15.0 | DataCenterBench Tier 1: Kimi K3 · Robocurve |
| Grok 4.7 | 13.3 | DataCenterBench Tier 1: Grok 4.7 · Robocurve |