CharXiv (descriptive and reasoning combined)
Multimodal · 2024-06-26
Pooled accuracy over CharXiv's 5,000 validation questions: 4,000 descriptive and 1,000 reasoning questions on 1,000 arXiv charts. It mixes the two splits that Atlas keeps as separate families, so it is a diagnostic.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 96.4 |
| Claude Sonnet 5.5 | 96.1 |
| GPT-6 Astra | 96.1 |
| GPT-6.1 Sol | 96.0 |
| Gemini 3.8 Flash | 95.4 |
| Gemini 3.7 Flash | 95.3 |
| Muse Spark 1.2 | 95.3 |
| Muse Spark 1.1 | 95.2 |
| Claude Fable 5.1 | 95.1 |
| Gemini 3.6 Flash | 95.1 |
| GPT-6 Sol | 95.0 |
| Claude Opus 5 | 94.6 |
| Gemini 3.1 Pro Preview | 94.6 |
| GPT-5.5 | 94.6 |
| Gemini 3.5 Flash | 94.3 |
| Claude Haiku 5.5 | 94.1 |
| GPT-5.6 Terra | 94.1 |
| Kimi K3 | 93.9 |
| Muse Spark 1.3 | 93.9 |
| GPT-5.4 | 93.4 |
| Opus 4.8 | 93.2 |
| GPT-5.6 Luna | 92.9 |
| Grok 4.6 | 92.9 |
| Grok 4.5 | 92.8 |
| Opus 4.7 | 92.7 |
| GLM-5.3 Flash | 92.5 |
| Sonnet 5 | 92.2 |
| GPT-5.6 Sol Pro | 92.1 |
| Grok 4.7 | 92.0 |
| Qwen3.5 397B A17B | 91.6 |
| GPT-6 Luna | 90.9 |
| Sonnet 4.6 | 90.5 |
| DeepSeek V4.1 Flash | 90.1 |
| Claude Fable 5 | 89.1 |
| Inkling | 88.9 |
| MiniMax M3 | 85.9 |
| GLM-5.3 | 63.0 |