The upgrade question sounds like it needs a feature table with scores in it. It does not. OpenAI’s launch page names exactly two measured differences between GPT Image 2.5 and the previous GPT-Image-2 generation, and then describes the rest of the improvement in words rather than numbers. What follows is that record, kept separate from the third-party benchmarks that fill the rest of the search results.
What OpenAI actually publishes
The two numbers on the record
Both appear on the same launch page, and both are relative rather than absolute.
- Image generation latency was “reduced by up to 50% compared with Images 2.0”
- Flare is documented as “delivering higher-quality images than GPT-Image-2 at 50% lower latency”, and as “the default choice for most applications”
Note what these are not. Neither is a benchmark score, neither names a task, and together they still do not say how fast anything runs on your workload.
The one number that is not OpenAI’s
A third figure circulates widely and comes from a customer rather than the vendor. Manus, quoted in the launch post, reports that Flare “delivers high-quality images at two to four times the speed of GPT-Image-2”. That is a partner’s measurement of its own workloads, not a published benchmark, and it should be read as a range a customer observed rather than a specification.
There is no gpt-image-2.5 model ID
This is the migration detail that breaks code first.
- The API publishes
GPT-Image-2.5 FlareandGPT-Image-2.5 Sunburst. There is no model calledgpt-image-2.5 - Both variants share one parameter set, so moving from the previous generation is a model-name change rather than an interface rewrite
- Choosing between the two is a separate question, worked through in the Flare and Sunburst comparison; the capability overview covers the shared parameters
| Dimension | Documented change | Not documented |
|---|---|---|
| Latency | Up to 50% lower than the previous generation | No per-request time figure |
| Quality | “Higher-quality images than GPT-Image-2” | No score, no benchmark |
| Editing scope | “Better at editing only what you’ve asked for” | No accuracy or success rate |
| Multi-turn behaviour | More reliable across multiple edits | No published turn limit |
| Price | None | No per-image price |
| Previous generation | Named as the comparison point | No sunset or deprecation date |
What changes in practice
The launch page describes the upgrade in capability terms. Four of them change how work feels, rather than how fast it runs.
- Scoped edits. The model is “better at editing only what you’ve asked for, while keeping the rest of the details the same—even with more complex subjects and backgrounds”.
- Multi-turn consistency. Across longer conversations it “follows specific editing instructions more reliably across multiple edits”, earlier changes “are more likely to stay consistent”, and quality should not degrade as edits accumulate.
- Fidelity to references. Generated results show “more natural lighting and richer textures” and are “better at preserving the subjects in your reference photos”.
- Complex layouts. “Images that include real-world information have more accurate content, and the model can handle more complex layouts including transparent backgrounds.”
The developer-facing framing is the most concrete of the four: the update lets an application “update a single element—such as a product, background, or piece of copy—while preserving the subject, composition, and brand treatment around it”. That is a statement about what a targeted edit may leave alone, which is easier to turn into an acceptance test than any quality adjective.
What OpenAI does not publish
The gaps matter more than usual on a migration page, because that is where teams guess.
- No per-image price. The launch post links to a separate pricing page instead of listing a rate.
- No quota or rate limit for either variant.
- No deprecation date, sunset window or migration deadline for the previous GPT-Image-2 generation.
- No benchmark comparing the two generations, and no quality score for either.
- No prompt migration guide, and no published list of prompt differences between them.
If a page gives you a date on which the previous generation stops working, or a dollar figure per image, check whether OpenAI published it. On the pages used here, it did not.
A migration checklist you can run
Four checks, in the order that catches the most breakage first.
Before you switch
1. Model name: replace the old ID with gpt-image-2.5-flare (default) or -sunburst.
2. Parameter diff: the two variants share one set; confirm nothing you pass was dropped.
3. Regression set: re-run the 10-20 prompts that matter, keep the acceptance criteria identical.
4. Cost: record tokens and response time before and after; the vendor publishes no per-image rate.A prompt worth keeping
Keep one prompt in the set that specifically tests the documented change, so the run measures the upgrade rather than the weather.
Start from a product photo on a plain background.
Change only the background to a workshop bench, and leave the product,
its shadow and its label exactly as they are.
Return one image, same framing, no extra objects, no text.Run it before and after the switch and compare how much of the frame moved. A model that is “better at editing only what you’ve asked for” should keep the label where it was.
When staying put is the sane call
Nothing OpenAI publishes forces the move. There is no sunset date on the previous generation, so the honest criteria are yours: whether your acceptance set passes on the current model, whether the latency change is worth a re-qualification cycle, and whether you need the scoped-edit behaviour the launch page describes. Teams whose acceptance criteria already pass and whose prompts are tuned have a real reason to wait, and there is no published deadline arguing against it.
Frequently asked questions
Is GPT Image 2.5 faster than GPT Image 2?
OpenAI documents latency “reduced by up to 50% compared with Images 2.0” and, for Flare, “50% lower latency” than GPT-Image-2. A customer quoted on the same page reports two to four times the speed. Treat the first two as the official record and the third as a partner measurement.
Do I need a new API key?
There is no published statement that credentials change with the model, so this is a confirmation step rather than an assumption. Check it on OpenAI’s own API documentation before planning a rollout.
Will my old prompts still work?
No prompt migration guide is published, and no list of prompt differences exists. The interface shape is close, so the documented change sits in behaviour rather than syntax; re-run your acceptance set rather than rewriting prompts speculatively. The what’s new summary lists the capabilities added in this generation.
Is the previous generation going away?
Nothing on the pages cited here sets a date. Any sunset date you see quoted elsewhere is someone else’s claim until it appears on OpenAI’s own model or deprecation page.
Which variant should I target first?
Flare, because it is documented as the default for most applications. Move individual workflows to Sunburst afterwards, where the edit precision is worth the longer generation time. Access and limits are covered in the free access guide.