TokenSwap-Bench (Image-Interleaved): leaderboard
Metric: Accuracy (%; 1,516 MMLU questions in which visualizable concepts (4.58 per question on average) are replaced by semantically aligned images inside the text, four-option multiple choice, standard prompt, answer letter extracted by regular expression). Source: arxiv.org. Saturation forecast: Estimated already saturated. 42 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 92.6 |
| 2 | Gemini 3.1 Pro (Preview) | 91.4 |
| 3 | Claude Sonnet 4.6 | 90.6 |
| 4 | GPT-5.1 | 90.3 |
| 5 | Gemini 2.5 Pro | 88.8 |
| 6 | GPT-5.2 | 88.7 |
| 7 | Gemini 2.5 Flash | 88.5 |
| 8 | Gemini 3 Flash | 88.4 |
| 9 | GPT-5 | 85.6 |
| 10 | Claude Haiku 4.5 | 83.3 |
| 11 | Qwen 3 VL 8B (Thinking) | 80.8 |
| 12 | Qwen 3 VL 4B (Thinking) | 78 |
| 13 | Gemini 2.5 Flash Lite | 76.6 |
| 14 | GPT-4.1 | 76.1 |
| 15 | Gemini 3.1 Flash Lite | 75.9 |
Interactive version: theaggregate.ai/benchmark?slug=tokenswap-bench-image-interleaved · How It Works · Data refreshed daily, snapshot 2026-09-29.