The RoboVista leaderboard has been updated (2026-10-15) and now covers 30 models (23 open-weight, 7 API) on 460 multiple-choice spatial-reasoning questions over real robot scenes, each run with a standard and (for most) a chain-of-thought prompt. The benchmark is available on Hugging Face.

Gemini 4 Argon is the new top model at 81.3%, ahead of GPT-6 Astra and Claude Fable 5.1, at the lowest evaluation cost of the three. Compared with Gemini 3 Flash, it improves in every robotics domain. (Announcement on X)

Model Weights Standard CoT Scene Planning
gemini-4-argon API 81.3 81.3 80.7 81.8
gpt-6-astra API 75.0 73.9 74.5 75.7
gemini-3-flash-preview API 68.9 65.4 69.9 67.6
claude-fable-5.1 API 63.9 — 65.0 60.8
qwen3.5-397b-a17b open 55.2 52.8 58.2 48.6
gpt-5-0 API 54.6 54.6 56.2 51.4
qwen3.6-27b open 53.5 54.4 55.9 48.0
gemma4-31b open 51.5 55.4 54.9 43.9

Top 8 by standard-prompt accuracy (%). The best open-weight model is qwen3.5-397b-a17b at 55.2%. See the full leaderboard for all 30 models and the per-domain breakdown.

Accuracy versus total eval cost: Gemini-4 Argon, GPT-6 Astra, Claude Fable 5.1
Gemini 4 Argon versus Gemini 3 Flash accuracy in each robotics domain
Charts from the announcement on X. They were made before the leaderboard's answer re-parsing, so their scores are a few tenths of a point higher (e.g. 81.9% vs. 81.3% for Gemini 4 Argon).