ChatGPT Images 2.5 was reported on September 8, 2026, with two new API model IDs: gpt-image-2.5-sunburst and gpt-image-2.5-flare. The announcement states that OpenAI's image-generation models have been used to create more than three billion images across ChatGPT Images and the GPT-Image API models. The version is described as improving instruction-following across multiple turns, responding faster, and better preserving subjects in reference photos.1

Context

The document demonstrates one API use: Willison upgraded his openai_image.py command-line tool to accept reference images and invoked it to add a raccoon scientist studying a chart, using the model gpt-image-2.5-sunburst with an external image URL. A resulting image description shows the chart with the cartoon added, providing exactly one concrete example of the claimed reference-image editing behavior.1

Assessment

What is new here is not merely an incremental image-model update but the productized split between editing-precision and everyday-generation behavior. The quoted guidance positions Sunburst for workloads where editing precision matters most and Flare for fast, high-quality everyday generation. If this distinction is meaningful, it changes the operational constraint for developers: the same API now offers an explicit trade-off between editing fidelity and generation speed, rather than forcing both goals through one model and a single prompt.1

Caveats

The supporting evidence is a single blog post with one user-visible edit, and the claims are inherited from OpenAI marketing language. What would change my mind is controlled third-party measurement: if multi-turn editing and reference-subject preservation show no meaningful improvement over prior models, or if Sunburst and Flare turn out to be behaviorally indistinguishable in independent tests, the announced shift is mostly narrative. Conversely, detailed API documentation, pricing, latency numbers, and systematic evaluations would strengthen the case that a real constraint moved.1