BEAKER - Trajectory Planning: leaderboard
Offline multimodal chemical-laboratory execution QA, with 1,000 expert-authored items across eleven tasks. Published task score (0–100). BEAKER v1, single-turn standardized QA; original-resolution images; short/mid videos 8/32 frames, long videos 128 frames in32grids; temperature0.7 when configurable. MCQ accuracy, affordance IoU, exp(-mean discrete Frechet distance/0.40) trajectory score, expert-semantic fill-in correctness, each scaled0–100. Overall uses expert weights0.25/0.15/0.35/0.25; experimentation image/video weights2:5. Source checkpoint/effort ambiguity is retained.
Metric: Published task score (0–100). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (Benchmark Sweep Run: source=BEAKER; single-turn; temperature=0.7 when configurable; image=original; video=8/32/128frames; checkpoint and effort as published) | 48.9 | |
| 2 | GPT-5.5 (Benchmark Sweep Run: source=BEAKER; single-turn; temperature=0.7 when configurable; image=original; video=8/32/128frames; checkpoint and effort as published) | 47.2 | |
| 3 | Claude Sonnet 4.6 (Benchmark Sweep Run: source=BEAKER; single-turn; temperature=0.7 when configurable; image=original; video=8/32/128frames; checkpoint and effort as published) | 46.3 | |
| 4 | GPT-5.4 (Benchmark Sweep Run: source=BEAKER; single-turn; temperature=0.7 when configurable; image=original; video=8/32/128frames; checkpoint and effort as published) | 46.1 | |
| 5 | Gemini 2.5 Pro (Benchmark Sweep Run: source=BEAKER; single-turn; temperature=0.7 when configurable; image=original; video=8/32/128frames; checkpoint and effort as published) | 44.4 | |
| 6 | Grok 4.20 (Reasoning) (Benchmark Sweep Run: source=BEAKER; single-turn; temperature=0.7 when configurable; image=original; video=8/32/128frames; checkpoint and effort as published) | 43.3 | |
| 7 | MiMo-V2-Omni (Benchmark Sweep Run: source=BEAKER; single-turn; temperature=0.7 when configurable; image=original; video=8/32/128frames; checkpoint and effort as published) | 42.6 | |
| 8 | Kimi K2.5 (Benchmark Sweep Run: source=BEAKER; single-turn; temperature=0.7 when configurable; image=original; video=8/32/128frames; checkpoint and effort as published) | 42.5 | |
| 9 | Kimi K2.6 (Benchmark Sweep Run: source=BEAKER; single-turn; temperature=0.7 when configurable; image=original; video=8/32/128frames; checkpoint and effort as published) | 41.8 | |
| 10 | Qwen 3.6 Plus (Benchmark Sweep Run: source=BEAKER; single-turn; temperature=0.7 when configurable; image=original; video=8/32/128frames; checkpoint and effort as published) | 41.1 |
Interactive version: theaggregate.ai/benchmark?slug=beaker-trajectory-planning · How It Works · Data refreshed daily, snapshot 2026-11-06.