Can AI observe and reason like a scientist?
ScienceVision evaluates whether models can obtain the visual evidence a scientific question needs directly from native 3D objects — molecular structures, crystal packings, CT volumes, CAD assemblies and building models — and then reason from it. The benchmark contains 250 expert-written questions across five disciplines and 22 subdomains, each paired with a structured reference answer and scored component by component.
Overall results
Values are the mean answer score on [0,1], with ineligible trials counted zero. 3D gives the agent the source artifact and code execution. 3D + MCP adds inspection tools modelled on expert workflow: view control, target isolation, slice browsing and measurement. 2D supplies six rendered views as a fixed-view control. Selected available repeats are averaged within each task, then tasks are weighted equally.
| Model | 3D | 3D + MCP | 2D · 6 views |
|---|---|---|---|
| GPT-6 Astra (Codex) | 0.5291 | 0.5282 | 0.2498 |
| GPT-6.1 Sol (Codex) | 0.4893 | 0.5110 | 0.2044 |
| Claude Fable 5.1 | 0.4619 | 0.4631 | 0.2965 |
| Claude Opus 5.5 | 0.4539 | 0.4343 | 0.2383 |
| Gemini 4 Argon | 0.4278 | 0.4321 | 0.1462 |
| Qwen3.8 Max 0902 | 0.3830 | 0.3736 | 0.2450 |
| Qwen3.8 Flash | 0.3616 | 0.3681 | 0.2468 |
| Kimi K3 | 0.3609 | 0.3369 | 0.2148 |
| GLM-5.3 Flash | 0.2846 | 0.2858 | 0.1521 |
| DeepSeek V4 Flash | 0.2652 | 0.2678 | 0.1661 |
| Doubao Seed Evolving | 0.2435 | 0.2527 | 0.1102 |
| Hunyuan-4 Preview | 0.1598 | 0.1738 | 0.0824 |
Harness. GPT-6 Astra and GPT-6.1 Sol were run in a Codex harness; the other ten models in Claude Code. A GPT row therefore measures a different scaffold, not only a different model. The two 3D settings average up to three runs per question; Gemini's carried-over segment has two runs. The 2D control uses three runs, with three Hunyuan-4 questions at two. Scores splice carried-over and re-run segments measured on different days, and model aliases carry no version.
The 2D column uses the same metric but a different evidence route. Compare its ordering with the 3D columns, not its absolute values.