GPT-Image-2.5

OpenAI shipped GPT-Image-2.5 on 8 September 2026. In ChatGPT it appears as ChatGPT Images 2.5, across the consumer, Work, and Codex tiers. In the API it arrives as two separate models.

Flare and Sunburst

gpt-image-2.5-flare is the default and the one most people should call. It generates at roughly half the latency of GPT-Image-2 at higher quality. gpt-image-2.5-sunburst is the precision model, aimed at edits where exact placement matters more than the wait. Both carry the same published token rates, so you’re choosing on speed against precision, not on budget. Dated snapshots of both are pinned to 2026-09-08.

What improved

Sharper detail, better fidelity when you hand it a reference image, and more consistent results when you edit the same image over several turns. The failure mode of conversational image editing has usually been the fourth edit quietly undoing the second, and 2.5 holds together better across a run of them.

What it costs

Image pricing is token-based rather than per-image, which makes budgeting awkward. The published rates are $5 per million text-input tokens, $8 per million image-input tokens, and $30 per million image-output tokens, with cached input cheaper. What any single picture costs depends on its dimensions, the quality setting, and how many reference images you fed in.

Where the upgrade actually shows

Not where you’d look for it first. GPT-Image-2 was already strong enough that side by side, 2.5’s output doesn’t announce itself. The difference I notice is in comprehension: hand it a rough description of an idea and it’s much better at working out what you were actually trying to visualize, and at picking a clever way to show it. That matters more than another notch of fidelity, because the slow part of making an image was never the rendering. It was the four attempts it took to explain what you wanted.

That keeps it the leader in most categories of image creation and editing, and close enough to flawless that I stopped keeping a shortlist of alternatives. It covers two jobs that usually need different tools: creative image work, and the unglamorous editing that goes into UI design. Midjourney is still the one to reach for when you want a particular look rather than a particular idea.

The Verdict

Best for: Creative image work and UI design alike: conversational editing, precise inpainting, and text-heavy visuals that have to stay legible.