HieroglyphBench
Multimodal · 2026-06-24
HieroglyphBench is a vision-language OCR benchmark for ancient Egyptian hieroglyphs. It contains 30 curated photographed inscription columns from the Pyramid of Unas, and models must output the ordered sequence of Gardiner sign-list codes; scores are sign accuracy from edit distance against expert transcriptions.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 4 Argon | 69.7 |
| Gemini 3.7 Flash | 55.0 |
| Gemini 3.8 Flash | 54.6 |
| Gemini 3.5 Flash | 52.5 |
| Gemini 3 Flash Preview | 48.8 |
| GPT-6 Astra | 40.0 |
| Gemini 3.1 Pro Preview | 39.7 |
| Claude Opus 5.5 | 37.4 |
| GPT-6.1 Sol | 31.4 |
| Muse Wonder | 29.7 |
| GPT-5.6 Sol | 27.6 |
| GPT-6 Sol | 26.0 |
| Kimi K3 | 25.9 |
| Muse Spark 1.1 | 24.5 |
| Claude Fable 5 | 22.9 |
| GPT-5.5 | 22.8 |
| Grok 4.5 | 21.2 |
| GPT-5.6 Terra | 20.5 |
| Claude Fable 5.1 | 20.4 |
| Grok 4.6 | 20.3 |
| Kimi K2.6 | 19.2 |
| Gemini 2.5 Pro | 18.8 |
| Grok 4.20 | 16.8 |
| GPT-5.4 Mini | 15.7 |
| Claude Opus 5 | 14.6 |
| Opus 4.8 | 14.0 |
| Inkling | 12.5 |
| GLM-5.3 Flash | 12.3 |
| DeepSeek V4.1 Flash | 12.3 |
| MiniMax M3 | 11.1 |
| Qwen3.7-Plus (source unspecified) | 10.1 |
| GPT-4o (source unspecified) | 9.5 |