OpenAI closes the computer vision gap in its GPT-5.6 family with a leap in detection alongside cost and stability challenges

Specialised benchmark tests conducted by computer vision platform Roboflow on OpenAI's new GPT-5.6 model family, which includes the Sol, Terra, and Luna models, have shown a tangible shift in computer vision and image processing capabilities, after the company's official announcement focused on computer use, user interface agents, and 3D rendering. Preliminary benchmarking results evaluating multimodal vision models revealed that the Sol model represents the strongest vision model OpenAI has offered to date, achieving a notable advance in object detection and counting tasks compared to the previous generation, GPT-5.5.

The object detection test recorded the clearest jump in the evaluation, with detection accuracy for the Sol model rising to 46.2 points under the mean average precision metric at fifty percent intersection over union (mAP@50), compared to around 13.8 points recorded by GPT-5.5. The Terra and Luna models ranked close behind, scoring 44.7 and 43.3 points respectively, moving object detection from a historic weakness for the company to a practically viable capability.Turning vision from a technical bottleneck into a practical tool begins with improving document layout analysis and detecting crowded spatial scenes with clear coordinate precision.This strength was evident in parsing complex document structures by distinguishing headings, paragraphs, tables, images, and signatures, alongside solid handling of scenes with adjacent items such as batches of eggs and medical pills, despite the challenge of models generating classifications and coordinates as text, which can raise error probabilities as the number of items increases.

The tests also revealed that model accuracy is linked to prompting style, as GPT-5.6 models achieved their highest accuracy levels when requesting absolute pixel coordinates in a specific format, whereas mismatched coordinate formats led to a loss of around fifteen points on the accuracy scale. Conversely, the evaluation observed instability where the Sol model returned bounding boxes that were random or arranged in straight lines that did not match reality when processing images of two thousand pixels or more. The OpenAI team confirmed this and attributed it to lower reasoning effort, making resizing or cropping images prior to submission the most practical workaround to avoid increases in latency and cost.

In visual counting tests, the accuracy of the Sol model rose to seventy-three percent compared to around sixty-four point nine percent for the previous generation, outperforming in tallying overlapping items such as metal brackets and bullet holes within specified zones of shooting targets. However, challenges were evident with blister packs of medication due to optical reflections and subtle differences between filled and empty pockets, alongside errors when counting non-standard types of confectionery. In optical character recognition and text extraction, the models maintained performance comparable to the previous generation, with Sol achieving an exact match accuracy of ninety point seven percent, successfully reading dirty tyre numbers and extracting live scores from sports broadcasts, while struggling with expiry dates printed vertically in low colour contrast.

These practical gains present organisations with a delicate balance between performance, cost, and processing latency, as Sol averaged around ten seconds per image at a cost of roughly two and a half cents, ranking second in cost behind Claude Fable 5.The engineering balance between speed and cost offered by Luna makes it the most practical option within the family, while external models remain competitive in data-heavy environments.Luna delivers latency of under five seconds and costs less than half a cent per image, while Gemini 3.5 Flash offers an alternative at eight-tenths of a cent while maintaining the lead in detection and counting tasks, requiring development teams to scrutinise deployment options based on data volume and latency requirements.