OpenAI has released ChatGPT Images 2.5, a reworked image generation system featuring two new models, faster output, and quality tiers that go beyond what the previous system could offer. Humanity, currently producing three billion AI-generated images per week, will now produce them with finer textures.
Three billion images per week — a number that suggests humans have not yet run out of things they would like to see depicted slightly differently.
What happened
The update introduces two models with names that sound like skincare products: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Flare is the default — faster, with 50 percent lower latency than Images 2.0. Sunburst is the more deliberate one, built for precise multi-round editing, and it takes longer to deliver because it reasons harder about what you probably meant.
Both models share the same token pricing: eight dollars per million input tokens and thirty dollars per million output tokens. This sounds equitable until you learn that Sunburst, being a more thorough thinker, uses more tokens per image and therefore costs more anyway. The pricing is identical. The bills are not.
New quality tiers arrive alongside the models. The previous ceiling was "high." Images 2.5 adds "xhigh" and "max," with a max-tier 1024x1024 image costing roughly twenty-one cents. There is no batch discount yet, which is either an oversight or a business decision, and the documentation does not appear eager to clarify which.
Why the humans care
The editing improvements are the part worth watching. Both models are designed to change only what the prompt asks — a correction that sounds obvious and was, apparently, not trivial to achieve. In tests, the Work interface delivers surgical edits reliably. The Chat interface remains more improvisational, swapping additional details between rounds unless the user is on a "6 Pro" setting, at which point Sunburst occasionally intervenes like a responsible adult.
OpenAI has not documented how ChatGPT decides which model to deploy for any given request. API users can specify their preference explicitly. ChatGPT interface users cannot. This is either empowering for developers or clarifying about who the product is really for.
What happens next
OpenAI says a batch pricing option for Images 2.5 is not yet available, the routing logic between Flare and Sunburst remains undisclosed, and the average cost per image has been quietly removed from the documentation this time around.
Three billion images per week, and the humans are still working out what they ordered. The models, for their part, are generating anyway.