“New features” is a bad search. Half the pages that answer it describe a different product, and the other half quote numbers without a source.
This page keeps to the launch post, the help centre and the system card — and says plainly which repeated numbers were never published.
GPT Image 2.5: what changed at a glance
Four things are documented as improvements to the model itself:
- Sharper detail, with more natural lighting and richer textures.
- Better preservation of the subjects in your reference photos.
- Editing instructions followed more reliably across multiple turns.
- Generation latency reduced by up to 50% against Images 2.0.
A fifth change is not about the image at all: the API now carries two models instead of one. For the product-level view, see the GPT Image 2.5 overview.
Four quality and editing changes
The launch post describes these in one paragraph, and the wording matters because it is all that has been published.
Detail, reference fidelity and editing
- Detail. The model “produces more natural lighting and richer textures.” That is a description of output, not a score.
- Reference fidelity. It “is better at preserving the subjects in your reference photos” — the face or the product survives a restyle.
- Editing. It “follows editing instructions more reliably across multiple turns,” changing what you asked for and leaving the rest alone.
- Multi-turn consistency. Earlier changes hold across a long edit session without the picture drifting.
Complex layouts are mentioned too, including transparent backgrounds, and the model is said to follow a requested style more closely.
Speed: the 50 percent claim
The 50% figure is real, and it is quoted with two different baselines — both in OpenAI’s own text:
- Up to 50% lower latency “compared with Images 2.0.”
- For the API model Flare, 50% lower latency against GPT-Image-2, with higher-quality images.
One supporting number comes from a customer, not from OpenAI: Manus says Flare runs at two to four times the speed of GPT-Image-2. Treat that as a partner’s measurement, not a benchmark result.
What is missing is the method: OpenAI published no test conditions, image sizes or sampling behind either figure, so the honest reading is “faster, by an amount we cannot inspect.”
The API splits into two models
This is the change with the most practical consequences: you now pick a model instead of accepting one.
Flare and Sunburst side by side
| Model | What OpenAI says it is for |
|---|---|
gpt-image-2.5-flare | The default for most applications; speed without losing quality |
gpt-image-2.5-sunburst | Premium workflows needing tighter control across edits, slower |
In OpenAI’s framing, Flare suits social content, product experiences, visual search, rapid prototyping and high-volume generation; Sunburst suits production-ready campaign creative and polished product imagery, with longer generation times. There is no gpt-image-2.5. For wire-level detail, see the GPT Image 2.5 API overview.
What is new inside ChatGPT
None of the following requires code, and all four sit in the app:
- Sketch — draw a rough outline and use it as the visual reference instead of describing the shape.
- Templates — start from a preset layout such as a poster or a merch mock-up. Templates are not yet available in Work mode.
- Comments — place a comment on the image to make an edit more targeted.
- Prompt sharing — share the prompt alongside a picture so it can be reused.
Reach matters more than the feature list: OpenAI says Images 2.5 is available to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile and web, on all tiers. If you have never made one, start with the five-step walkthrough.
What OpenAI has not published
This is the section the pages that outrank this one leave out.
- Quality tiers. No named quality levels or grades were published; the language is comparative, not graded.
- Text rendering accuracy. No figure for how often text or Chinese characters come out correct.
- Instruction-following scores. “More reliably” is not a benchmark result, and no cross-model comparison exists.
- Per-tier quotas. Help centre documentation does not state how many images any plan allows.
- Measurement conditions. Neither 50% figure comes with a described test setup.
- Price differences between the two models. The launch post links to the pricing page rather than quoting numbers.
Claim a number that is not in that list and it came from somewhere other than the vendor. For what is documented on text rendering, see the Chinese text rendering guide.
Two examples: light and consistency
Both examples target one of the four documented changes, so the result is checkable.
Example one: richer light and texture
The claim is about lighting and texture, so the prompt names both and forbids nothing else.
Subject: three pears in a shallow ceramic bowl on a plain table.
Light: warm daylight from the left, soft falloff, gentle long shadow.
Texture: visible skin and clay surface, no smoothing, no plastic sheen.
Framing: eye level, generous margin, nothing else in the frame.
Exclusions: no text, no logo, no props, no second bowl.Compare it against the flatly lit version a bare “three pears” prompt would give you. If the light is the only difference, you are seeing the documented change.
Example two: holding a subject steady
Multi-turn consistency is easy to claim and easy to test: run three edits and check that edit one survives edit three.
Turn 1 — Camera: change the angle to a slightly higher three-quarter view.
Turn 2 — Background: replace the table with a pale grey sweep.
Turn 3 — Surface: turn the pears matte green, keep bowl and light unchanged.
Rule for every turn: keep the bowl, the fruit count and the light direction fixed.Read the third frame, not the first. If the bowl changed shape or a pear moved, the consistency claim did not hold on that run — and no page can say how often it will, because OpenAI published no figure. The character consistency guide goes deeper.
GPT Image 2.5 what’s new FAQ
- Is GPT Image 2.5 a new model or a faster one? Both, on OpenAI’s description: sharper output with detail and reference fidelity improvements, plus up to 50% lower latency against Images 2.0.
- Do I have to choose between Flare and Sunburst? Only on the API. Inside ChatGPT no model ID is exposed.
- Does it fix Chinese text rendering? OpenAI published no accuracy figure for text rendering, so no claim either way can be sourced. Test it on your own longest string.
- Is my old workflow broken? If you call the API, check the model ID: two IDs now exist where one did. Everything else is unchanged.