SteelBench - Macro-F1: leaderboard
SteelBench (2026) evaluates vision-language models on 1,345 real steel-plant CCTV clips with distant workers, dust, steam, glare and occlusion, using a 25-class action taxonomy and plant safety rules. Macro-F1 over the 25 action classes, in percent, higher is better.
Metric: Macro-F1 over the 25 action classes (%; 1,345 densely annotated clips of real steel-plant CCTV footage (8 frames at 1080p per clip) from six industrial domains, zero-shot structured prompt with the 25-class action taxonomy, site context, PPE definitions and safety rules; provider-default temperature; failed parses count as wrong). Source: arxiv.org. Saturation forecast: Around 2035. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 122B A10B | 22.6 | #331 |
| 2 | GPT-4o | 22.3 | #741 |
| 3 | Gemini 2.5 Pro | 17.8 | #266 |
| 4 | Claude Opus 4.7 | 17.8 | #79 |
| 5 | Gemini 2.5 Flash | 17.3 | #453 |
| 6 | Nemotron Nano 12B v2 VL | 17 | #2231 |
| 7 | Gemma 4 31B (IT) | 16.2 | #342 |
| 8 | GPT-5.4 | 16.1 | #123 |
| 9 | Llama 4 Maverick | 12.7 | #1099 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=steelbench-macro-f1 · How It Works · Data refreshed daily, snapshot 2026-11-13.