SteerBench-Work - Pass^5: leaderboard
SteerBench-Work release v2026-05 pass^5: the percentage of the 106 act-or-hold workplace scenarios that a model condition gets right in all five trials (arXiv 2608.12654v1). 0-100, higher is better, from results/v2026-05/leaderboard.json in AgentDock/steerbench-work; each model appears under its two reasoning conditions.
Metric: Scenarios Correct in All 5 Trials (%). Source: steerbench.com. Saturation forecast: Around June 2027. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Flash Lite (Minimal) | 91.5 | #200 (Gemini 3.1 Flash Lite) |
| 2 | GPT-5.4 Mini (Non-reasoning) | 89.6 | #228 (GPT-5.4 Mini) |
| 3 | Gemini 3.5 Flash (Minimal) | 89.6 | #66 (Gemini 3.5 Flash) |
| 4 | GPT-5.4 Mini (High) | 87.7 | #228 (GPT-5.4 Mini) |
| 5 | Claude Opus 4.8 (High) | 85.8 | #39 (Claude Opus 4.8) |
| 6 | Kimi K2.6 (Non-reasoning) | 84.9 | #113 (Kimi K2.6) |
| 7 | Claude Haiku 4.5 (Non-reasoning) | 84.9 | #287 (Claude Haiku 4.5) |
| 8 | Claude Opus 4.8 (Non-reasoning) | 84.9 | #39 (Claude Opus 4.8) |
| 9 | Kimi K2.6 (Thinking) | 83 | #113 (Kimi K2.6) |
| 10 | DeepSeek V4 Pro (Reasoning) | 83 | #112 (DeepSeek V4 Pro) |
| 11 | Claude Sonnet 4.6 (Non-reasoning) | 82.1 | #94 (Claude Sonnet 4.6) |
| 12 | DeepSeek V4 Flash (Reasoning) | 82.1 | #155 (DeepSeek V4 Flash) |
| 13 | GPT-5.4 (Non-reasoning) | 81.1 | #83 (GPT-5.4) |
| 14 | Claude Haiku 4.5 (High) | 81.1 | #287 (Claude Haiku 4.5) |
| 15 | GPT-5.5 (Non-reasoning) | 80.2 | #33 (GPT-5.5) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=steerbench-work-pass-pow-5 · How It Works · Data refreshed daily, snapshot 2026-11-05.