Pointerbench — Text
Multimodal · 2026-07-02
Pointerbench-Text is a 500-example synthetic GUI-grounding benchmark for words, characters, punctuation, caret positions, interface text, text bounding boxes, and invoice fields. Point tasks use point-in-bounding-box accuracy; bounding-box tasks require at least 90% target coverage and 70% prediction precision.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 58.4 |
| Sonnet 4.6 | 47.8 |
| Opus 4.8 | 44.4 |
| Opus 4.7 | 36.2 |
| GPT-5.6 Sol Pro | 32.0 |
| GPT-5.5 | 31.4 |
| GPT-5.6 Sol | 30.0 |
| GPT-5.4 | 25.0 |
| GPT-5.6 Terra | 22.0 |
| GPT-5.6 Luna | 20.0 |
| GPT-5.6 Terra Pro | 20.0 |
| GPT-5.6 Luna Pro | 18.0 |
| GPT-5 | 13.6 |
| Kimi K2.5 | 3.4 |
| Kimi K2.6 | 2.8 |
| Kimi K2.7 Code | 1.8 |
| GPT-5 Mini | 1.4 |
| Qwen3.6-Flash-2026-04-16 | 1.2 |
| Gemini 3.5 Flash | 0.6 |
| Qwen3.6 Plus (2026-04-02) | 0.6 |
| Qwen3.7-Plus | 0.6 |
| Gemini 3.1 Pro Preview | 0.4 |
| MiniMax M3 | 0.4 |
| Qwen3 VL 235B A22B Thinking | 0.4 |
| GPT-5 Nano | 0.2 |
| Qwen3.5-9B | 0.2 |
| Grok 4.3 | 0.0 |
| Grok Build 0.1 | 0.0 |
| MiMo-V2.5 | 0.0 |
| Step 3.7 Flash | 0.0 |