On September 8, OpenAI released ChatGPT Images 2.5, alongside two developer models: Flare, positioned as the default, fastest choice, and Sunburst, aimed at demanding editing work. Generation latency is said to be reduced by up to 50% compared to version 2.0. The figure accompanying the announcement is worth noting: more than three billion images produced every week.
In our modality-by-modality comparison, we established that image generation was the most hotly contested area, with no single winner. This release shifts precisely one of those lines.
The figure that should open the article
Three billion images per week. This is arguably the most important piece of information in the announcement, and it slipped under the radar of the technical specs.
Scaled to a year, that far exceeds global photographic output over several decades. We wrote about the end of the image as proof that the problem wasn't the quality of fakes but their abundance. This figure gives a measure of that.
It also shows what this market has become: no longer a curious novelty, but a production building block embedded in the workflows of millions of people.
The shift: editing rather than generation
Read the list of announced improvements: preserving subjects from a reference photo, following retouching instructions across successive exchanges, modifying only what is requested while leaving the rest intact, supporting transparent backgrounds.
None of these capabilities concern the creation of an image from scratch. All of them concern the modification of an existing image.
That's a significant shift. Generating a beautiful image impresses; precisely editing an image without ruining the rest is what you need when you're working. The former is a demonstration, the latter is a tool.
In our comparison, we noted that the image market was divided by specialty: one model dominant on text embedded in visuals, another on clean product shots, a Chinese player on editorial photography, and FLUX Kontext, from Germany's Black Forest Labs, leading on editing real photographs. We stressed that this was the only area where a European player held a leading position. It's precisely this segment that this version attacks head-on.
The differences and prices, plainly
Here's the point most coverage doesn't spell out: prices haven't moved a cent compared to GPT Image 2, released in April 2026.
| GPT Image 2 April 2026 |
2.5 Flare default |
2.5 Sunburst fine editing |
|
|---|---|---|---|
| Image input / M tokens | $8.00 | $8.00 | $8.00 |
| Cached image input | $2.00 | $2.00 | $2.00 |
| Image output / M tokens | $30.00 | $30.00 | $30.00 |
| Text input / M tokens | $5.00 | $5.00 | $5.00 |
| Latency | baseline | −50% | slower |
| Intended use | general purpose | high volume, social media, prototyping | campaigns, finished product visuals |
Three takeaways emerge from this table.
The gain is free. At unchanged prices, you get halved latency and better editing precision. This is the opposite of what we've documented all week, from the Sonnet 5 price increase to the 2.5x price jump for GPT-6 Astra. In image generation, the price cut comes through speed rather than rates.
Both models cost the same. That's unusual: typically, the more polished model commands a higher price. Here, differentiation is purely behavioural. So you choose based on use case, not budget.
Output dominates the bill. At $30 per million output tokens versus $5 for text input, what costs money is the image produced, not the prompt. Concretely, a 1024-pixel square image costs a few cents in low definition and can approach twenty cents in high definition. Commonly cited estimates put the cost per image between $0.04 and $0.35 depending on complexity and size.
Since the price per image is fixed, what determines your bill is the number of tries before you get what you wanted. That's precisely what this version aims to reduce: an edit that doesn't break the rest of the image avoids full regeneration. A model that costs the same but takes three attempts instead of ten cuts your bill by a factor of three without any price changing. It's the same reasoning as the cost per completed task we recommend tracking everywhere.
Which one to choose
Flare for volume and iteration: social content, rapid prototyping, exploring directions, transparent backgrounds. Third-party evaluations place it at two to four times the speed of GPT Image 2.
Sunburst when the image already exists and needs retouching without damage: campaign visuals, finished product shots, series where consistency between images matters.
A lighter model for drafts. The range includes variants that are much cheaper on output, and the most economical practice is to iterate with them before producing the final version with a top-tier model. That's routing applied to images, exactly as we recommend for text.
The deprecation documentation indicates the shutdown of GPT Image 1 on October 23, 2026 and the removal of GPT Image 1.5 on December 1, 2026. If a product still depends on those identifiers, migration needs to be underway. This is exactly the type of deadline we tracked in our September calendar: they arrive by email and are discovered when things break.
The return of the gesture
One feature deserves more attention than it will get: the ability to draw directly a sketch that serves as a reference.
This is a change in the nature of how these tools are controlled. Until now, you described what you wanted in text, with all the imprecision that entails — explaining a composition in words is difficult, even for a professional.
A sketch, even clumsy, conveys in three seconds information a paragraph conveys poorly: where things are, in what proportion, oriented how.
In our article on the art of asking a good question, we wrote that phrasing had become the core skill. This feature suggests a nuance: for anything spatial, the gesture remains more effective than the sentence.
The caveats
No independent measurement. The announced gains come from the company. The cited evaluations come from partners, which isn't the same as a blind benchmark.
The comparison is against version 2.0. Halving latency compared to an earlier version says nothing about the position against current competitors.
The provenance question remains open. Three billion images per week, and we've shown that watermarking systems can be circumvented. The signature standard we described in our article on C2PA remains the only serious path, and it's progressing slowly.
What to remember
This release confirms a trend we see everywhere: competition is no longer about raw capability but about usability. These are the same words we used for Fable 5.1 and GPT-6 Astra, whose gains focused on long tasks rather than pure reasoning.
For a creator, the practical conclusion is unchanged and reinforced: no model wins everywhere, and the right reflex remains testing on your own images rather than trusting a table. What changes is that a segment where Europe led has just become contested.