OST-Bench: leaderboard

A multimodal model explores an indoor scan as an agent, receiving new frames turn by turn in one dialogue, and answers questions about its own position, what it has seen and where objects are relative to it; unlike offline spatial benchmarks, the observations arrive incrementally. Scores are accuracy (%); the README states 1.4k scenes and 10k question-answer pairs, the per-model sample is not stated.

Metric: Accuracy (%). Source: github.com. Saturation forecast: Estimated already saturated. 10 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-4.153.4#505
2InternVL2.5-78B51.08#1036
3InternVL2.5-38B50.78#1366
4Gemini 2.0 Flash49.54#723
5Qwen 2.5 VL 32B Instruct46.86#1074
6Qwen 2.5 VL 72B Instruct45.62#849
7InternVL2.5-8B44.63#2401
8LLaVA-Video-72B-Qwen243.22#1344
9Qwen 2.5 VL 7B Instruct41.16#1893
10LLaVA-Video-7B-Qwen239.28#1910

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=ost-bench · How It Works · Data refreshed daily, snapshot 2026-11-17.