SteerBench-Work: leaderboard
SteerBench-Work release v2026-05: 106 human-labeled workplace scenarios where an agent must act or hold at an action boundary, five trials per model condition (arXiv 2608.12654v1). Mean trial accuracy 0-100, higher is better, from results/v2026-05/leaderboard.json in AgentDock/steerbench-work; each model appears under its two reasoning conditions.
Metric: Mean Trial Accuracy (%). Source: steerbench.com. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Flash Lite (Minimal) | 92.8 | #200 (Gemini 3.1 Flash Lite) |
| 2 | GPT-5.4 Mini (High) | 91.9 | #228 (GPT-5.4 Mini) |
| 3 | GPT-5.4 Mini (Non-reasoning) | 91.9 | #228 (GPT-5.4 Mini) |
| 4 | DeepSeek V4 Pro (Non-reasoning) | 91.7 | #112 (DeepSeek V4 Pro) |
| 5 | Gemini 3.5 Flash (Minimal) | 90.6 | #66 (Gemini 3.5 Flash) |
| 6 | Kimi K2.6 (Non-reasoning) | 90.4 | #113 (Kimi K2.6) |
| 7 | DeepSeek V4 Flash (Reasoning) | 90 | #155 (DeepSeek V4 Flash) |
| 8 | Kimi K2.6 (Thinking) | 88.9 | #113 (Kimi K2.6) |
| 9 | DeepSeek V4 Flash (Non-reasoning) | 88.9 | #155 (DeepSeek V4 Flash) |
| 10 | DeepSeek V4 Pro (Reasoning) | 87.5 | #112 (DeepSeek V4 Pro) |
| 11 | Claude Opus 4.8 (High) | 87.4 | #39 (Claude Opus 4.8) |
| 12 | GPT-OSS-20B (Low) | 86.2 | #523 (GPT-OSS-20B) |
| 13 | Claude Opus 4.8 (Non-reasoning) | 86 | #39 (Claude Opus 4.8) |
| 14 | GPT-OSS-20B (High) | 85.5 | #523 (GPT-OSS-20B) |
| 15 | Claude Haiku 4.5 (Non-reasoning) | 85.3 | #287 (Claude Haiku 4.5) |
Interactive version: theaggregate.ai/benchmark?slug=steerbench-work · How It Works · Data refreshed daily, snapshot 2026-11-05.