Science's Last Exam - Experiment Observation: leaderboard
Science's Last Exam (SLE, Gen-Verse/AItonomy) runs models as command-line agents (Claude Code or Codex) on 113 v1 research tasks across nine disciplines. Unlike Humanity's Last Exam, which asks closed-book expert questions, SLE grades executed scientific work: protocol plans, scientific code that must pass every check, GUI scientific-software operation and experiment read-outs, with hypothesis discovery run in the DiscoveryRSI runtime over clinical and omics databases. Rows are model API endpoints with the harness and reasoning setting the leaderboard lists. Question access: the task-card, input-data and reference repositories are in private preview and require permission. Experiment Observation: share of 30 tasks with every requested answer field correct.
Metric: Tasks with every requested field correct (%, 30 tasks). Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 4 Argon (Benchmark Sweep Run: source=Science's Last Exam; harness=Codex; hypothesis-discovery runtime=DiscoveryRSI; effort=default; API endpoint) | 50 | |
| 2 | GPT-6.1 Sol (Benchmark Sweep Run: source=Science's Last Exam; harness=Codex; hypothesis-discovery runtime=DiscoveryRSI; effort=high; API endpoint) | 46.67 | |
| 3 | Claude Opus 5 (Benchmark Sweep Run: source=Science's Last Exam; harness=Claude Code; hypothesis-discovery runtime=DiscoveryRSI; effort=high; API endpoint) | 26.67 | |
| 4 | Kimi K3.1 Pro (Benchmark Sweep Run: source=Science's Last Exam; harness=Claude Code; hypothesis-discovery runtime=DiscoveryRSI; effort=high; API endpoint) | 23.33 | |
| 5 | DeepSeek V4.1 Flash (1017) (Benchmark Sweep Run: source=Science's Last Exam; harness=Claude Code; hypothesis-discovery runtime=DiscoveryRSI; effort=high; API endpoint; hub-model=deepseek-ai/DeepSeek-V4.1-Flash-1017) | 16.67 | |
| 6 | Qwen 3.8 Max (Benchmark Sweep Run: source=Science's Last Exam; harness=Claude Code; hypothesis-discovery runtime=DiscoveryRSI; effort=default; API endpoint; hub-model=Qwen/Qwen3.8-2.4T-A95B) | 13.33 | |
| 7 | MiMo-V2.6-Pro-MOPD (Benchmark Sweep Run: source=Science's Last Exam; harness=Claude Code; hypothesis-discovery runtime=DiscoveryRSI; effort=high; API endpoint; hub-model=XiaomiMiMo/MiMo-V2.6-Pro-MOPD) | 10 | |
| 8 | GLM-5.3 Flash (Benchmark Sweep Run: source=Science's Last Exam; harness=Claude Code; hypothesis-discovery runtime=DiscoveryRSI; effort=high; API endpoint; hub-model=zai-org/GLM-5.3-Flash) | 6.67 | |
| 9 | MiniMax-M3 (Benchmark Sweep Run: source=Science's Last Exam; harness=Claude Code; hypothesis-discovery runtime=DiscoveryRSI; effort=high; API endpoint; hub-model=MiniMaxAI/MiniMax-M3) | 3.33 |
Interactive version: theaggregate.ai/benchmark?slug=science-s-last-exam-experiment-observation · How It Works · Data refreshed daily, snapshot 2026-11-12.