VisualNeedle: leaderboard
Frontier multimodal large language models (MLLMs) have been reported to achieve over 90% accuracy on fine-grained perception benchmarks. However, such scores do not necessarily imply faithful use of visual evidence. Prior studies have identified three shortcuts that inflate benchmark performance. First, linguistic priors and lexical cues in questions often enable models to infer plausible answers
Metric: w/ Tools Acc. (%) (self-reported). Source: benchmarklist.com. Saturation forecast: Around July 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 56.01 | #58 |
| 2 | Gemini 3 Flash | 48 | #99 |
| 3 | Seed 2.0 Pro | 36.8 | #118 |
| 4 | Seed 1.8 | 31.78 | #139 |
| 5 | GPT-5.4 | 31.23 | #79 |
| 6 | GPT-5.2 | 28.31 | #110 |
| 7 | Qwen 3.5 Plus | 22.11 | #125 |
| 8 | Kimi K2.5 | 19.32 | #144 |
| 9 | Qwen 3 VL 235B A22B (Thinking) | 11.44 | #238 (Qwen 3 VL 235B A22B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=visualneedle · How It Works · Data refreshed daily, snapshot 2026-10-20.