LG-Bench - Medicine: leaderboard
LG-Bench (2026): main Table 2 accuracy on 10,200 life-science multiple-choice questions: Medicine 3210, Biology 3630, Chemistry 3360. Identical prompt/evaluation protocol; source model labels do not bind base versus instruction checkpoints or API snapshots, so study identities remain separate. The independently reviewed inconsistent Llama-3.1-70B row is omitted pending correction, not repaired or replaced. GCS/KBS graph diagnostics are not accuracy and are not included.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LG-Bench Study GPT-5 | 91.31 | |
| 2 | LG-Bench Study OpenAI o3 | 90 | |
| 3 | LG-Bench Study OpenAI o4-mini | 89.75 | |
| 4 | LG-Bench Study Claude-3.5-sonnet | 88.91 | |
| 5 | LG-Bench Study Grok-2 | 88.54 | |
| 6 | LG-Bench Study GPT-4o | 88.47 | |
| 7 | LG-Bench Study Claude-3.7-sonnet | 88.41 | |
| 8 | LG-Bench Study Llama-3.1-405b | 88.07 | |
| 9 | LG-Bench Study Deepseek-v3 | 87.17 | |
| 10 | LG-Bench Study Grok-3 | 86.29 |
Interactive version: theaggregate.ai/benchmark?slug=lg-bench-medicine · How It Works · Data refreshed daily, snapshot 2026-10-21.