Atlas

Benchmarks

← All benchmarks

Pointerbench — Pro

Multimodal · 2026-07-02

Pointerbench-Pro is a 500-example synthetic GUI-grounding benchmark for icon, text, and mixed targets across 100 professional applications. A predicted absolute-pixel click is correct when it lands inside the target bounding box.

Top models (higher is better)

ModelScore
Claude Fable 580.8
Opus 4.880.2
Sonnet 4.677.4
GPT-5.575.8
Opus 4.775.4
GPT-5.6 Sol74.0
GPT-5.6 Terra Pro74.0
GPT-5.6 Sol Pro72.0
GPT-5.6 Terra70.0
GPT-5.6 Luna Pro62.0
GPT-5.6 Luna60.0
GPT-5.450.4
GPT-520.2
GPT-5 Mini5.8
MiniMax M33.6
Qwen3.6 Plus (2026-04-02)3.6
Qwen3 VL 235B A22B Thinking3.2
Kimi K2.53.0
Qwen3.7-Plus3.0
Kimi K2.62.8
Qwen3.5-9B2.6
Qwen3.6-Flash-2026-04-162.6
GPT-5 Nano1.6
Kimi K2.7 Code0.8
Gemini 3.5 Flash0.6
Gemini 3.1 Pro Preview0.4
Step 3.7 Flash0.2
Grok 4.30.0
Grok Build 0.10.0
MiMo-V2.50.0
Loading Atlas data…