According to CoinMeta, OpenAI has fixed an image encoding issue in GPT-6, SOL, GPT-6, and Luna. This issue reduced the model's ability to understand images, but after the fix, the visual tasks performed by API and Codex improved. Tests published by OpenAI showed that the average score of GPT-6 and Luna (within max) in the RefCOCOG visual localization evaluation increased from 28.8% to 60.0%, a 31.2 percentage point improvement. This test mainly assessed whether the models could accurately identify objects or regions in images based on descriptions. Both models were officially launched on API, and the update log on OpenAI recorded this fix on September 25th. OpenAI also recommends that developers using image inputs re-evaluate their workflows and try again those that were affected previously.