OpenAI introduced ChatGPT Images 2.5 on September 8, 2026, adding creative controls and two developer models: Flare and Sunburst. The announcement emphasizes reference fidelity, focused editing and faster generation. For a creative workflow, the decisive test is whether an edit changes the requested detail while preserving everything the reviewer already approved.
Flare and Sunburst target different iteration needs
OpenAI positions Flare for fast, high-quality everyday generation and Sunburst for more precise creative work with longer generation times. The product update also includes sketch references, image comments and reusable prompts. These features give a reviewer more ways to express a correction than a broad text instruction alone.
Evaluate model choice against the cost of rejection. A fast draft may be enough to select a composition; a detailed product edit may require stronger preservation of labels, proportions and background elements. Comparing generation latency without counting rejected outputs can favor a model that creates more total work.
Choose the integration route around editing continuity
The developer guide distinguishes direct Image API generation and edits from image generation inside Responses conversations. The latter supports multi-turn editing workflows. The identifiers are gpt-image-2.5-flare and gpt-image-2.5-sunburst; developers should follow the current guide for the route they use.
A one-shot asset generator has different needs from an editor that carries several revisions forward. Preserve the approved reference and the instruction for each edit. If the application silently uses an earlier image or drops a constraint, even a capable model can return an unsuitable result.
Test preservation as carefully as the requested change
Choose a reference with details that would be costly to lose: a face, a readable package label or a consistent room layout. Ask for one bounded change and inspect both the altered region and the supposedly preserved areas. Repeat this over a short sequence of revisions rather than judge only the first turn.
When an output is rejected, identify the failure: incorrect typography, identity drift, unintended object changes or composition. That record makes a model comparison actionable. A general judgment that one image is prettier does not show whether the workflow meets the brief.
Make delivery requirements explicit
Set dimensions, format, transparency and quality based on the destination before generating the final asset. Review the file at the scale where it will be used, including small text and cropped layouts. These checks concern the delivered image rather than the model's preview.
The release broadens creative capability, but it does not guarantee factual accuracy, brand compliance or rights to supplied material. Nerova has not independently reproduced the launch's quality or latency claims. The appropriate adoption test is a complete, reviewed revision task with its total cost and failure rate recorded.