According to tests conducted by the third-party organization Roboflow, the GPT-5.6 series has demonstrated a marked improvement in visual capabilities. Notably, Sol’s object detection score soared from 13.8 points in GPT-5.5 to an impressive 46.2 points. Terra and Luna also delivered commendable performances, showcasing enhanced counting accuracy and excelling in tasks such as document layout recognition and dense scene counting. However, Sol has experienced a regression compared to GPT-5.5 in the ability to extract text information directionally. Moreover, it exhibits a detection box offset issue in large-sized images, which can be mitigated by increasing the inference level or reducing the image size. In terms of cost, Sol is priced at approximately 2.5 cents per image, Terra at about 1 cent per image, and Luna at less than 0.5 cents per image. Meanwhile, Gemini 3.5 Flash offers even greater cost-effectiveness. The competitive landscape among visual large models has shifted from the question of "can they understand" to "how accurately they can perform tasks," and the race for cost-efficiency is now underway.