OpenAI's ChatGPT Images 2.5: Two Models, One Real Lesson for Builders

OpenAI shipped ChatGPT Images 2.5 this week, and the detail worth pausing on isn't the image quality bump — it's that OpenAI didn't ship one model. It shipped two: gpt-image-2.5-sunburst and gpt-image-2.5-flare, according to Simon Willison's writeup of the official announcement.

Per OpenAI's own API docs, Sunburst is positioned for "workflows where editing precision matters most," while Flare is for "fast, high-quality everyday image generation." Willison reads Sunburst as the stronger default for most serious use — which tracks: precision and consistency are what break first when you try to put image generation into a real product loop, not raw output quality.

Why the two-model split matters more than the upgrade

For anyone building on top of these APIs rather than just chatting with them, model-tiering isn't new — it's exactly the pattern OpenAI, Anthropic, and others already run for text: a frontier model for the cases that need it, a cheaper/faster sibling for everything else. Applying that same split to image generation signals OpenAI expects image APIs to get embedded into pipelines the same way text APIs have — batch jobs, agent loops, product features — where cost and latency per call actually matter at volume.

That's reinforced by the scale OpenAI is quoting: more than 3 billion images generated across ChatGPT Images and the GPT-Image API models to date. At that volume, a single one-size-fits-all model stops making economic sense — you want a fast, cheap option for the 90% of everyday requests and a precision option you reach for deliberately.

The capability gains that actually matter for agentic workflows

Two specific improvements stand out for builders rather than casual users:

• Better multi-turn instruction-following — the model holds context across a sequence of edit requests instead of drifting, which is the difference between a usable iterative editing loop and one that needs a fresh prompt every step.
• Better subject preservation from reference photos — the model keeps a person, product, or object consistent across edits instead of subtly re-drawing it each time.

Both of these are the exact failure modes that make image generation hard to wire into an agent: an agent that's iterating on a design or a product photo needs the model to remember what it's editing, not just what it's being asked to do. Willison's own openai_image.py CLI tool update — adding support for passing one or more reference images — is a small but telling sign of where the tooling ecosystem is already heading.

The builder takeaway

If you're integrating image generation into a product — a design tool, an e-commerce catalog pipeline, a marketing asset generator — the news here isn't "the images look better." It's that you now have an explicit precision-vs-speed dial in the API itself, and picking the wrong side of it will show up in your latency and your bill before it shows up in output quality. Treat the model choice as a deliberate architecture decision, the same way you'd choose between a reasoning model and a fast model for a text task.

Choose Sunburst for workflows where editing precision matters most, and Flare for fast, high-quality everyday image generation.

— OpenAI API documentation