Trainee-Bench: leaderboard
Workplace agent benchmark from The Agent's First Day (Fu et al., 2026): the model works as an intern in the authors' EvoEnv environment through 50 scenarios that each stream 2 to 6 rule-generated tasks with deadlines, time-critical interruptions and clues that must be found by exploring files or asking simulated colleagues. The score is the Checkpoint Score, the mean over scenarios of the fraction of verifiable task checkpoints passed, in percent; higher is better.
Metric: Checkpoint Score (%; mean share of passed checkpoints per scenario). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5 (Thinking) | 72 | #256 (GLM-5) |
| 2 | MiniMax-M2.5 | 65 | #590 |
| 3 | Gemini 3 Flash | 63 | #159 |
| 4 | Claude Sonnet 4 | 59 | #397 |
| 5 | Grok 4 | 54 | #306 |
| 6 | GPT-5.1 | 49 | #225 |
| 7 | GPT-4o | 38 | #741 |
| 8 | Qwen 3 VL 235B A22B | 37 | #477 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=trainee-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.