Sample images generated with ChatGPT Images 2.5, from OpenAI’s announcement post. Credit: OpenAI OpenAI released ChatGPT Images 2.5 with two new API models, one of which Adobe has already put into Firefly. The company says more than 3 billion images a week are created across ChatGPT and its image API. The notable additions are interface features rather than model research, including the ability to share the prompt alongside an image OpenAI released ChatGPT Images 2.5 on Tuesday. The company said in its announcement that the model brings sharper detail, more precise editing and up to 50% lower latency than Images 2.0.
Two new models go to developers. GPT-Image-2.5 Flare is the default, and Sunburst is the slower, higher-control option for production creative work. Adobe is a launch customer The detail worth pausing on is in the testimonials. Adobe says the new GPT-Image-2.5 models are now in Firefly, its creative AI studio.
Matt Chotin, a senior director of product at Adobe, framed Firefly as bringing leading AI models together with Adobe’s own tools in one place. That is a company positioning itself as the interface rather than only the model. It is a deliberate strategy, not a concession. Firefly has been built as a conversational layer over the whole toolset, and model choice is part of that pitch. The traffic runs both ways Adobe is inside ChatGPT too. More than 70 Adobe tools are available there.
So each company is a distribution channel for the other. Neither can be confident it owns the moment a creative task begins. That ambiguity is the whole competitive question in creative software right now. Whoever owns the starting point captures the workflow.
Prompts become objects OpenAI is answering that question with a social feature. When you share an image now, you can attach the prompt that made it, and someone else can run it with their own photos. OpenAI links to a viral 1980s headshot prompt as the example. That is a distribution mechanic dressed as a convenience.
A shareable prompt is a template that spreads on its own. It also keeps the loop inside ChatGPT rather than inside anyone else’s canvas. The other new features point the same way Sketch lets you draw in ChatGPT and use the drawing as a reference. Templates offer starting formats such as posters and merchandise, and comments can be placed directly on an image.
None of this is model research. It is interface work, and interface work is what you do when you want people to stay. Availability is broad. Images 2.5 is rolling out to ChatGPT, ChatGPT Work and Codex users on every tier, across desktop, mobile and web.
Three billion a week The scale figure is the one to sit with. OpenAI says more than 3 billion images are created weekly across ChatGPT Images and the GPT-Image models in the API. That is the company’s own number and is not independently audited. Taken at face value it is a vast quantity of synthetic imagery entering circulation every week.
OpenAI says it applies C2PA metadata and invisible watermarking, alongside prompt and image checks, and has published a system card for the release. Provenance marking at that volume is a genuine engineering commitment. Whether provenance survives contact Marking images is necessary and not sufficient. Metadata is routinely stripped by platforms, screenshots and re-encoding, and invisible watermarks vary in robustness.
Regulators are still working out what any of it means legally. Brazil has been defining what counts as a deepfake ahead of its vote. The prompt-sharing feature cuts against provenance in one respect. It makes it easier for a distinctive look to be reproduced by many people from the same recipe. The competitive picture The reference-photo improvements target the most commercially valuable use.
Putting a real person or product into new settings is what advertising, ecommerce and social content actually need. Higgsfield and Manus, both quoted in the announcement, are building on the API rather than competing with it. Manus says Flare runs two to four times faster than the previous model in its own evaluations. Others are going the other way. Freepik rebranded as Magnific as a profitable AI creative platform on $230m of recurring revenue, built as its own destination.
What to watch Watch whether Adobe keeps naming its model suppliers. A studio that aggregates models is strong while it can choose, and weaker if one supplier becomes indispensable. Watch the prompt-sharing numbers too, if OpenAI ever publishes them. Adobe has been assembling its own answer across Firefly, GenStudio and partnerships, and the contest is over where work starts rather than which model renders it. Also tagged with
