GPT-5.6 Sol: OpenAI Finally Steps Up Its Computer Vision

Last week OpenAI unveiled the GPT-5.6 lineup with the Sol, Terra and Luna models, putting the spotlight on “computer use” — agents navigating desktop applications and working with 3D visualizations. All of this rests on the quality of visual perception, so the Roboflow team ran the newcomers through its soon-to-be-released benchmark for multimodal models. It covers object detection, counting, OCR and data extraction.

Detection and counting

The most noticeable jump is in object detection. The previous GPT-5.5 scored just 13.8 mAP@50, whereas Sol reached 46.2; Terra and Luna are close behind (44.7 and 43.3). Detection went from an outright weak spot to a practically usable feature. Sol confidently parses document structure — headings, paragraphs, tables, images and captions — and does a decent job with dense scenes such as a scatter of pills or eggs, where identical objects are easy to miss or count twice.

There are quirks, too: on images of roughly 2000×2000 pixels and up, Sol sometimes places boxes at random spots, arranging them in neat rows that bear no relation to the actual objects. OpenAI confirmed that at a low “reasoning level” the model loses stability on large pictures; the fix is either raising the reasoning effort (at the cost of tokens and latency) or downscaling the image beforehand.

Object counting improved across the whole lineup: Sol delivered 73.0% versus GPT-5.5's 64.9%. Even the cheapest model, Luna, surpassed OpenAI's previous level. Sol managed to count overlapping metal staples and tallied bullet holes only within the specified zones of a target, demonstrating an understanding not just of “what to count” but also of “where.”

OCR and the price question

Text recognition, on the other hand, stayed at its former level — 90.7% versus GPT-5.5's 91.2% — and in pinpoint data extraction Sol actually fell behind its predecessor. The model reads handwritten notes and lettering on the curved surface of a dirty tire well, but stumbles on small vertical text with glare — for instance, an expiration date on a blister pack.

The progress came at a cost in resources. Sol spends about 10 seconds per image and roughly 2.5 cents — making it the second most expensive model after Claude Fable 5. Terra and Luna are faster and cheaper. Gemini 3.5 Flash, meanwhile, costs 0.8 cents and, according to the authors, still leads in detection and counting, remaining the more economical choice for high-volume tasks.

The overall takeaway: OpenAI has taken vision seriously and has closed much of the gap with the leaders, but unconditional first place is still a way off.

Source: Hacker News