ScienceArena (Vision Routes): leaderboard

ScienceArena (Zhao et al., 2026) evaluates language models on 13 recent science olympiad exams in physics, chemistry and biology, graded with the official process-credit rubrics by an LLM judge calibrated against medalist grading. The score is the mean of the 13 exam scores normalized to 0 to 100; this board holds vision-capable routes, which see the original question pages of the four 2026 contests, and higher is better.

Metric: Weighted average (0-100): mean of the 13 contest scores, each normalized to 0-100 by its full score, over IBO 2023, APhO, EuPhO, USNCO, IPhO, IChO, CPhO and CChO 2025 and INChO, NBPhO, USAPhO, IPhO and IChO 2026; interleaved solving, one subpart at a time with earlier turns in context; open-ended physics and chemistry turns graded against the official solution and process-credit rubric by the authors' medalist-calibrated LLM-judge pipeline; the nine other contests use text-only question renderings for every model and the four 2026 contests (NBPhO, USAPhO, IPhO and IChO 2026) are solved from the original question-page images (vision-capable routes). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.

Top models

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=sciencearena-vision-routes · How It Works · Data refreshed daily, snapshot 2026-11-13.