ChatGPT Images 2.5 is a meaningful upgrade if your work continues after the first generation. It improves reference fidelity, targeted editing and consistency across multiple revisions, while OpenAI reports generation latency up to 50% lower than Images 2.0. ChatGPT also gains Sketch, Templates, image comments and shareable prompts. The API now offers two models: fast, everyday GPT-Image-2.5 Flare and precision-focused GPT-Image-2.5 Sunburst.
The launch looks strongest as an editing release, not merely a prettier text-to-image model. Early tests from Axios, TechRadar and Forbes broadly support that reading, but their samples are limited. This review separates those observations from OpenAI's specifications, pricing and launch claims.
ChatGPT Images 2.5 at a glance
| Item | What is available | What to know |
|---|---|---|
| Release date | September 8, 2026 | Facts and links checked September 10, 2026 |
| ChatGPT access | All ChatGPT tiers, ChatGPT Work and Codex | Desktop, mobile and web; feature rollout can vary |
| Core model changes | Sharper detail, improved reference fidelity, precise edits and multi-turn consistency | These are OpenAI's launch claims; early reviews are encouraging but not benchmarks |
| Reported speed | Up to 50% lower latency than Images 2.0 | A maximum vendor-reported reduction, not a guarantee for every prompt |
| New ChatGPT tools | Sketch, Templates, comments and prompt sharing | Templates are not yet available in Work mode |
| API models | gpt-image-2.5-flare and gpt-image-2.5-sunburst | Both accept text and image inputs and return images |
| API endpoints | Images API and the Responses API image-generation tool | Image edits are supported by Sunburst; Flare's model page lists generations |
| Provenance | C2PA metadata and invisible watermarking | Useful signals, but publishing teams still need disclosure and review policies |
The source of record is . Product-specific exceptions come from the current and .
What actually changed in the image model?
Images 2.5 targets the parts of creative work that usually force a restart: preserving a subject from a reference, changing one area without disturbing the rest, and carrying approved decisions through several revisions. That makes the update more relevant to campaign assets, product images, portraits and layout exploration than a one-prompt beauty contest.
Better reference fidelity
OpenAI says the model preserves recognizable subjects and distinctive features more reliably while changing setting, style or composition. Lighting is intended to look more natural and textures richer. The practical benefit is not perfect identity lock; it is a better chance that a person, pet, object or room still reads as the same subject after a transformation.
Input image from .
OpenAI's output changes the clothing while retaining the pose, framing and blue background. This is a curated official example, not an independent benchmark.
The pair shows what “reference fidelity” means more clearly than a general quality claim. The outfit changes, but the pose, framing and background remain anchored. Production teams should still test varied skin tones, fine accessories, logos, hands and less controlled source photos before relying on this behavior at scale.
More focused edits
Images 2.5 is designed to change the requested element while leaving approved composition and brand treatment intact. OpenAI demonstrates changes to products, backgrounds and copy without rebuilding the entire scene. For a creative team, fewer accidental changes can matter more than higher detail because every unwanted drift creates another review cycle.
This remains generative editing rather than pixel-locked retouching. Check faces, product geometry, labels, dates, prices and legal copy after every revision. Keep the original source asset and approved versions outside the chat so a convincing edit cannot quietly replace a correct one.
Stronger multi-turn consistency
The model is also intended to preserve earlier changes through longer conversations. In practical terms, “make the jacket green,” followed by “remove the sign,” should be less likely to undo the jacket or degrade the subject. That improves iteration, but a long chat is not a version-control system. Name each approved state, export it, and compare the next revision against that checkpoint.
Faster generation, with a careful qualifier
OpenAI reports up to 50% lower image-generation latency than Images 2.0. “Up to” matters: complexity, quality setting, resolution, load and the chosen API model can change total wait time. The useful consequence is a faster feedback loop, not a universal promise that every image completes in half the time.
Sketch, Templates and comments change how you prompt
The new product features reduce how much spatial intent must be squeezed into prose.
Sketch lets mobile users draw a rough arrangement and then describe the finished image. A sketch can communicate placement, scale and negative space faster than a paragraph. It is especially useful for poster layouts, product compositions, room concepts and presentation graphics.
Templates provide structured starting points for common outputs such as flyers, logos and product photography. ChatGPT can ask follow-up questions before generating. That turns a vague request into a lightweight brief, although OpenAI notes that Templates are not yet available in ChatGPT Work mode.
Comments place instructions on specific parts of an image. The official release notes describe commenting on a generated image to request focused changes. TechRadar's reviewer found the point-and-comment workflow particularly useful because it removes ambiguity about which object should change.
Prompt sharing lets another person reuse the creative recipe with their own details or photo. It can speed up internal handoff, but a shared prompt alone does not preserve every hidden product setting or guarantee identical output. Record the model, date, quality, aspect ratio and source assets alongside the prompt.
What independent hands-on reviews found
The first independent reports agree on the direction of travel: editing and control are the headline gains. Their evidence is useful as early observation, not as a leaderboard.
a cat transformation, a tattoo extension and a logo-to-merchandise workflow. The reviewer reported stronger likeness preservation and found sketching plus visual edit targeting compelling. The first logo attempt was underwhelming and improved after another direction, a useful reminder that the model still benefits from art direction.
and highlighted the Edit toolbar, Sketch and Templates. The article also notes that some toolbar features appeared before the formal launch, so the complete interface change should not be attributed solely to the model. Its verdict is one experienced user's workflow assessment, not a controlled comparison.
. The model asked clarifying questions, offered visual directions, produced a cover in under 45 seconds in that run and retained core details during a later style change. That is a useful example of guided ideation for a non-designer, but one timed generation is not a latency distribution.
Together, these reports support a restrained conclusion: Images 2.5 appears easier to steer and revise than its predecessor. They do not establish error rates for text, identity preservation or large production batches. Teams should run a fixed prompt set with blind review, revision counts, completion time and usable-output rate before switching a pipeline.
GPT-Image-2.5 Flare vs Sunburst
Developers get two API models with different operating priorities.
| Model | Best fit | Official positioning | Trade-off |
|---|---|---|---|
gpt-image-2.5-flare | Everyday generation, previews, high-volume iteration | OpenAI's fastest high-quality image model | Choose only after checking whether its edit support and fidelity meet your workflow |
gpt-image-2.5-sunburst | Detailed product imagery, campaign creative and precise editing | OpenAI's most capable image generation and editing model | Longer generation time than Flare |
Both models accept text and image inputs, support low, medium, high, xhigh, max and auto quality settings, and can be selected in the Image API or through the Responses API image-generation tool. OpenAI lists stable and dated IDs, including gpt-image-2.5-sunburst-2026-09-08 and gpt-image-2.5-flare-2026-09-08. Pin a snapshot when visual behavior must remain stable across a campaign.
See the official model pages for and . Start with Flare for exploration, then route approved or edit-sensitive work to Sunburst. Do not assume the expensive-looking name is always necessary; measure usable outputs per dollar and per minute.
GPT Image 2.5 API pricing
OpenAI bills both 2.5 models at the same published token rates.
| Meter | Flare | Sunburst |
|---|---|---|
| Text input, per 1M tokens | $5.00 | $5.00 |
| Image input, per 1M tokens | $8.00 | $8.00 |
| Cached image input, per 1M tokens | $2.00 | $2.00 |
| Image output, per 1M tokens | $30.00 | $30.00 |
Those are the current rates, verified September 10, 2026. The listed per-token rates are twice GPT Image 2's $2.50 text input, $4 image input, $1 cached image input and $15 image output rates.
That does not prove every 2.5 image costs exactly twice as much. Per-image cost depends on the tokens consumed by the request and output. OpenAI's Sunburst model page explicitly says the GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption. Until a 2.5 calculator or usage study is available, budget from actual API usage rather than a guessed cents-per-image figure.
A practical workflow for better results
- Start with the deliverable. State the intended asset, audience, aspect ratio and where copy must remain readable.
- Attach authoritative references. Use original product, person or location images and identify which details must not change.
- Use a sketch for layout. Show placement and negative space when prose becomes awkward.
- Approve one state. Export the strongest result before asking for another edit.
- Make one focused change per turn. Point or comment on the target and repeat the invariants.
- Finish critical text outside the pixels. Names, prices, dates and legal lines belong in an editable layout layer.
- Record provenance. Save prompt, model ID, source assets, generation date and human approvals.
For a broader production process, Mengbi's connects the brief, generation and final checks. Our compares ChatGPT with tools that may offer stronger visual exploration, typography or design-system controls.
Safety, rights and disclosure still need a workflow
OpenAI says Images 2.5 continues to use prompt and image safeguards, C2PA metadata and invisible watermarking. Its also warns that stronger realism can increase deepfake risk and describes checks at input and output layers.
These controls do not replace permission, trademark review or human disclosure. Get consent before editing identifiable people, verify the rights for every reference, and do not treat an AI rendering of a product, place or event as documentary evidence. Exporting, resizing or publishing through another tool can also affect visible provenance signals, so keep an internal asset record.
Who should use ChatGPT Images 2.5?
Upgrade now if you already use ChatGPT Images for iterative work, frequently start from reference photos, or lose time because small edits disturb approved details. Flare is a sensible API starting point for previews and routine content; Sunburst is the better candidate when precision edits determine whether an asset ships.
Test before migrating if your output includes exact packaging, long embedded copy, strict character identity or regulated claims. Build a representative evaluation set and compare failure rate, revision count, time and total API cost against Images 2.0 and your current alternative.
Keep specialist tools in the loop when you need layers, vector output, typography systems, asset libraries, approvals or pixel-level retouching. Images 2.5 can shorten the path from idea to strong draft. It does not turn a chat into a complete design-production stack.
Frequently asked questions
Is ChatGPT Images 2.5 available to free users?
Yes. OpenAI says Images 2.5 is rolling out across all ChatGPT tiers on desktop, mobile and web. Existing image-generation limits still apply, and individual features may appear at different times.
What is the difference between Flare and Sunburst?
Flare prioritizes fast, high-quality everyday generation. Sunburst prioritizes editing precision and detailed creative work, with longer generation times. Both have the same published token rates, so choose using quality, latency and usable-output tests rather than price alone.
Is ChatGPT Images 2.5 always 50% faster?
No. OpenAI reports latency reductions of up to 50% compared with Images 2.0. That is a maximum first-party figure. Prompt complexity, model, quality, resolution and service load can change the result.
Does Images 2.5 fix text inside images?
OpenAI says the model better understands complex layouts and real-world information, while early reviewers report useful results. Important text still needs character-by-character review. Put long copy, prices, dates and legal language into an editable design layer for final production.
Can Images 2.5 edit an existing photo?
Yes. ChatGPT accepts uploaded images and natural-language edits. Sunburst supports the Images API edit endpoint. Get permission for identifiable subjects and keep the unedited original.
How much does one GPT Image 2.5 image cost?
OpenAI publishes token rates, not one fixed price per image. Both Flare and Sunburst list $5 per million text-input tokens, $8 per million image-input tokens and $30 per million image-output tokens. Actual image cost depends on token consumption, and the existing calculator does not yet estimate 2.5 usage.
Are Sketch and Templates available everywhere?
Sketch is documented for mobile, while image creation and editing are available on web, iOS and Android. OpenAI says Templates are not yet available in ChatGPT Work mode. Rollout and interface details can vary by platform.
Verdict
ChatGPT Images 2.5 improves the part of image generation that most affects real work: keeping a good idea intact while you refine it. The combination of reference fidelity, focused edits, multi-turn consistency and spatial controls makes it a strong general-purpose creative tool. Flare and Sunburst also give API teams a clearer speed-versus-precision choice.
The caveats are equally practical. Early third-party tests are small, generative edits are not pixel-locked, and current token rates do not translate into a trustworthy fixed cost per image. Treat the release as a promising production candidate, then validate it with your own references, failure cases and invoice.
Last updated: September 10, 2026. Product availability and API pricing can change; follow the linked official sources for current terms.