GPT 5.6 Sol: OpenAI's Best Vision Model to Date

OpenAI's GPT-5.6 lineup (Sol, Terra, Luna) targets powerful visual understanding capabilities for computer use and UI agent implementation.

Object Detection

• The Sol model saw a dramatic improvement in object detection performance, reaching a practical level with 46.2 mAP@50 compared to 13.8 mAP@50 in GPT-5.5.
• It excels at detecting document layouts (titles, paragraphs, tables, signatures, etc.), making it suitable for data extraction workflows.
• To achieve optimal results, prompts should be configured to return absolute XYXY coordinates based on image pixels.
• Stability may decrease with large images exceeding 2,000 x 2,000 pixels, so resizing or cropping is recommended.

Object Counting

• The Sol model scored 73.0%, showing improved performance over GPT-5.5 (64.9%).
• It demonstrated excellent comprehension in high-difficulty tasks such as identifying overlapping objects or counting only objects within specific zones.

OCR and Data Extraction

• OCR performance remained similar to GPT-5.5, but its ability to read text and follow instructions in complex visual scenes, such as text on curved surfaces or scores on sports broadcast screens, is outstanding.