VisualNeedle - Without Tools: leaderboard

VisualNeedle (2026) tests active visual search on 300 questions over information-dense high-resolution scenes whose decisive evidence is a small region not readable at a glance. This board is accuracy in percent when the model answers from the full image without the crop tool, mean of three runs; higher is better.

Metric: Accuracy (%) on the 300 VisualNeedle questions when the model answers from the full image without the Crop tool, mean of three runs at the generation settings of Table 4; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Seed 1.8High17.89#252
2Gemini 3.1 Pro (Preview)High16.34#99
3Gemini 3 FlashHigh15.12#175
4Qwen 3.5 PlusThinking15#242
5GPT-5.4xHigh14.32#135
6Seed 2.0 ProHigh13.11#222
7Kimi K2.5Thinking10.78#266
8GPT-5.2xHigh10.56#185
9Qwen 3 VL 235B A22B (Thinking)8.34#459

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

Interactive version: theaggregate.ai/benchmark?slug=visualneedle-without-tools · How It Works · Data refreshed daily, snapshot 2026-11-21.