OpenAI GPT-5.6 Sol Vision Capabilities and Benchmarks

GPT-5.6 Sol advances OpenAI's visual understanding

OpenAI's GPT-5.6 lineup—comprising Sol, Terra, and Luna—marks a substantial leap in vision capabilities, particularly in object detection and counting. While GPT-5.6 Sol is the most capable vision model OpenAI has released to date, it faces stiff competition from Gemini 3.5 Flash in terms of cost, latency, and raw detection accuracy.

Significant Gains in Object Detection and Counting

GPT-5.6 Sol transforms object detection from a major weakness into a practical capability. In Roboflow's benchmarks, Sol achieved a 46.2 mAP@50 score, a massive increase from GPT-5.5's 13.8. Terra and Luna followed closely with 44.7 and 43.3 respectively.

Key Detection Strengths

• Document Layout: Sol excels at identifying titles, paragraphs, tables, images, and signatures, facilitating document workflows.
• Dense Scenes: The model handles scenes with many similar, closely packed objects (e.g., pills or eggs) more effectively than previous iterations.
• Coordinate Prompting: Performance is highly sensitive to prompt formatting. GPT-5.6 models perform best when prompted for absolute XYXY coordinates in image pixels; using incorrect formats can reduce performance by approximately 15 mAP points.

Counting Accuracy

Counting performance improved across the entire lineup. Sol reached 73.0% accuracy, up from 64.9% for GPT-5.5. The model demonstrated an ability to count overlapping objects and apply counting rules to specific zones (e.g., counting bullet holes only within scoring zones).

OCR and Data Extraction Performance

OCR performance remained relatively stable compared to GPT-5.5. Sol achieved a 90.7% mean similarity score, slightly behind GPT-5.5's 91.2%. In text extraction, Sol scored 82.5%, while GPT-5.5 scored 87.6%.

Despite the slight dip in raw scores, Sol demonstrated high utility in complex scenarios:

• Handwritten Text: Successfully produced full transcriptions and extracted specific dates from handwritten notes.
• Embedded Text: Read tire size sequences on curved, dirty surfaces and extracted live scores from sports broadcasts.
• Failure Points: Small, vertical, low-contrast text affected by reflections (such as expiration dates on blister packs) remains a challenge.

Operational Trade-offs: Cost, Latency, and Stability

Improvements in vision capabilities come with increased token usage and higher operational costs.

Model	Avg. Time per Image	Avg. Cost per Image
GPT-5.6 Sol	~10 seconds	~$0.025
GPT-5.6 Terra	~6 seconds	~$0.01
GPT-5.6 Luna	~5 seconds	<$0.005
Gemini 3.5 Flash	N/A	~$0.008

Stability Issues

OpenAI has confirmed that Sol becomes less stable on images 2,000 x 2,000 pixels or larger. While higher reasoning effort can improve stability, it increases cost and latency. The recommended workaround is to resize or crop images before API submission.

Community Insights and Counterpoints

While the benchmark data shows progress, community discussion highlights several critical perspectives on the practical application of these models:

Competitive Positioning

Several users noted that Gemini 3.5 Flash often outperforms Sol in detection and counting while being significantly cheaper. One user pointed out:

GPT 5.6 Sol was outperformed on all benchmarks by Gemini 3.5 Flash, apart from a single exception (OCR) where Fable was the winner. Gemini 3.5 Flash not only outperformed GPT 5.6 Sol, but did so at 1/3 of the cost.

Specialized Use Cases

Users reported success in niche areas where general benchmarks may not capture value:

• UI/UX Analysis: One user found Sol superior to Claude for identifying non-normative UI blocks and restructuring pages into composable units.
• Video Captioning: A user claimed Sol is currently the best video captioning model, specifically for accurately captioning complex sub-second movements.
• Specialized Transcription: A user reported that Sol successfully transcribed sheet music, a task that typically trips up most vision models.

Technical Skepticism

Some developers argued that many of the showcased tasks (like counting pills) are better handled by traditional computer vision tools. As one user noted:

Ironically, the pill counting example selected to showcase "the best vision model" can be easily solved with OpenCV template matching, a technology created 25 years ago.