OpenAI has officially launched ChatGPT Images 2.5, introducing two distinct models designed to balance speed and precision for the more than three billion images users generate weekly. The update splits the workload between GPT-Image-2.5 Flare, which serves as the default for most tasks, and GPT-Image-2.5 Sunburst, a heavier model built for complex visual work requiring tighter control. Flare delivers a significant performance boost, cutting generation latency by up to 50 percent compared to the previous Images 2.0 while maintaining higher quality. This split allows developers to choose explicitly via the API, though ChatGPT’s interface currently lacks clear controls for selecting between the two.
The core improvement lies in editing stability. Previous versions often altered unrelated parts of an image when users requested minor tweaks. Images 2.5 aims to change only the specified elements, preserving the rest of the composition across multiple rounds of edits. In tests using ChatGPT Work with GPT-6 Astra, the model successfully adjusted specific details like a banana’s color without distorting the background or other subjects. However, results in standard Chat mode remained inconsistent, with unintended changes persisting unless the “6 Pro” setting was active. This suggests that while the underlying technology is sharper, its application varies significantly depending on the user’s subscription tier and interface mode.
Pricing remains complex despite OpenAI’s claim that both models share the same token rates: $8 per million input tokens and $30 per million output tokens. The actual cost per image diverges because Sunburst typically requires longer reasoning runs, making it more expensive than Flare even at identical token prices. New quality tiers labeled “xhigh” and “max” have been introduced, with the “max” tier costing roughly $0.21 per 1024x1024 image. Notably, OpenAI has removed the cheaper batch rate available in previous versions, leaving users without a low-cost option for bulk generation. The lack of transparency regarding how ChatGPT routes requests between Flare and Sunburst adds another layer of uncertainty for casual users expecting consistent performance.
