Adobe wired OpenAI's Images 2.5 into Firefly on launch day

A design lead with twelve seats to renew can now reach the same OpenAI renderer from two different subscriptions, and the renewal turns on which one her team opens first thing on Monday.

Teams often assume model quality decides which tool survives a renewal, but the purchase actually decides where the file lives afterwards. Adobe's own framing, from senior product director Matt Chotin, is that Firefly brings leading AI models together with Adobe's tools in one place ^[6]. That is an aggregator's pitch, and it holds for exactly as long as Adobe can change the supplier without users noticing.

The evidence that the model layer has become interchangeable is thinner than the framing suggests. It is one launch testimonial ^[4] plus a pitch in which model choice is a feature ^[7]. The reporting does not say whether a Firefly user picks between the two new models or whether Adobe assigns one ^[8]. Developers calling the API do get the choice: Flare is the default, and Sunburst is the slower, higher-control option for production creative work ^[2].

Every performance number here is self-reported. OpenAI claims sharper detail, more precise editing and up to 50% lower latency than Images 2.0 ^[3]. Manus, quoted in the same announcement, says Flare runs two to four times faster than the previous model in its own evaluations ^[9]. Those are different measurements from parties with something to sell.

The scale figure is volume: more than 3 billion images a week across ChatGPT and the image API, OpenAI's own count and not independently audited ^[10]. Divide by seven and that is about 429 million a day, or roughly 4,960 images every second ^[11]. It is the only usage disclosure in the release; no retention split accompanies it, and no measure of how far a working session goes ^[12].

Against that volume OpenAI applies C2PA metadata and invisible watermarking, runs prompt and image checks, and has published a system card ^[13]. Thenextweb.com's read is that this is necessary and not sufficient, because metadata gets stripped in routine handling by platforms, screenshots and re-encoding, while invisible watermark robustness varies ^[14]. The new sharing feature pulls the other way. A shared image can now carry the prompt that made it, and the recipient can run that recipe on their own photos, with OpenAI pointing to a viral 1980s headshot prompt as the example ^[15]. TNW argues that makes one distinctive look reproducible by many hands from a single source ^[16]. The rest of the release is interface work too: sketching as a reference, templates for posters and merchandise, comments placed on the image itself ^[19].

Aggregation is one bet. Freepik went the other way, rebranding as Magnific and running as its own destination, profitable on $230m of recurring revenue ^[17].

For a specific team, two axes settle it. Where does the task start, in a chat window or on a canvas. Where must the artifact land, in a versioned file or in a post. A task that begins in chat and ends as a post makes the canvas seat decoration. One that starts on the canvas and ends in a versioned file makes the model's name a procurement footnote. Work that crosses from chat into a versioned file is where Friday's pain sits, because a human moves that asset and its provenance travels badly. Sort last quarter's jobs into those four boxes; the fullest box names the seat that survives the renewal.