DataCenterBench
Robocurve, Inc.Maintainer ↗10 single steps from data-center manuals, controls (knob, key, switch, cover, slider, keypad) and connectors (plug, unplug, Ethernet, cable clip).
As of November 17, 2026, GPT-6 Astra leads with 40.0, followed by GPT-6.1 Sol (38.3) and Gemini 4 Argon (35.0). Official leaderboard ↗
Leaderboard
- Tasks
- 10
- Attempts / task
- 3
- Median cost / task
- $1.52
- Median time / task
- 8 min 29 s
Overall vs speed
Published time per task attempt
Overall vs cost
Published cost per task attempt
Score 0 to 100, the mean of three tries per task, judged from photos.
GPT-6 Astra40.083.310033.3$19.123 min 35 s2
GPT-6.1 Sol38.383.310033.3$10.473 min 03 s3
Gemini 4 Argon35.01000.0066.7$1.8331 min 17 s4
Claude Opus 5.521.750.033.366.7$17.6714 min 52 s5
Muse Wonder 1.016.733.333.30.00$1.528 min 29 s6
GLM 5.3 Flash15.016.70.0033.3$0.01416 min 16 s7
Kimi K315.050.00.000.00$0.60420 min 30 s8
Grok 4.713.333.366.70.00$0.5965 min 05 s9
DeepSeek V4.1 Flash11.716.70.000.00$0.1387 min 05 sTasks
10 tasks. Hover a task to play its clip (6 of 10).
▶ Clip
▶ Clip

▶ Clip
▶ Clip
▶ Clip
▶ Clip
Methodology
- Scoring
- Staged rubric
- Time cap / attempt
- —
- Control rate
- —
- Models in index
- 9 of 9
Scoring. Judge scores each trial 0, 0.5 or 1 against written criteria.
Who runs it. Independent evaluator (Robocurve), LLM judges with a human tie-break.
Robots. I2RT YAM (two-arm).
Note. Coding agents only. Same rigs and operator as RoboDojo-RC, three tries per task.
Time per attempt. Published or measured for 9 of 9 models.
Cost per attempt. Published for 9 of 9 models.