GPT Image 2.5: what actually changed

At a glance

• It's two models, not one. Flare is the fast tier, which OpenAI says runs at 50% lower latency than GPT Image 2. Sunburst is the precision tier and takes longer. A bare gpt-image-2.5 isn't a valid model string.

• Editing is the real story. OpenAI's system card says users can change an image's setting, style or composition while keeping more of the original detail. That's the failure mode this release targets.

• The grain is a carryover, not a regression. Users have reported noise artifacts since GPT Image 2.0, and 2.5 didn't fix them. TechRadar tested four images and sided with the complaints, while calling the effect far from a disaster.

• Sketch, templates and comments belong to ChatGPT. The API gives you the models and their parameters, not the interface built on top of them.

• Check your dependencies. Neither 2.5 model supports the Batch API, and gpt-image-1 shuts down on 2026-10-23, with gpt-image-1.5 and gpt-image-1-mini following on 2026-12-01.

What OpenAI shipped on September 8

OpenAI released GPT Image 2.5 on 8 September 2026. The consumer product is called ChatGPT Images 2.5. In the API it arrived as two separate models, which is the first thing worth knowing because most write-ups treat it as a single release.

The improvement everyone will notice is in editing. Generating a good image from a text prompt stopped being the hard part a while ago; the hard part was asking for one change and getting back a picture that had quietly redrawn the parts you liked. That is what 2.5 is built to fix, and it's what OpenAI's own materials lead with.

Everything below comes from OpenAI's documentation, system card, changelog and announcement post. The model is two days old, so there's no controlled benchmark yet; where users and reviewers have reported problems, we've linked them.

Flare and Sunburst do different jobs

The two variants split on speed against precision. Flare is the everyday model, described in the docs as "our fastest model for high-quality, everyday image generation". The 50% figure comes from OpenAI's announcement post, which calls Flare "the default choice for most applications, delivering higher-quality images than GPT-Image-2 at 50% lower latency". Sunburst is the docs' "most capable model for image generation and editing", and the announcement pairs its extra precision with longer generation times.

Both accept the same quality settings: low, medium, high, xhigh, max and auto, and both bill at GPT Image 2 token rates. They can still cost different amounts per image, and the image generation guide is blunt about why: "Equal token rates don't mean equal cost per image: token consumption can differ by model and quality setting."

One practical trap for anyone writing against the API: there is no gpt-image-2.5 model identifier, and that URL returns a 404. You name Flare or Sunburst explicitly. Snapshot pinning works the usual way, with gpt-image-2.5-flare-2026-09-08 and its Sunburst equivalent.

Editing that leaves the rest of the frame alone

The system card says the model "produces more consistent results when editing images, and improves infographic accuracy and layout", and that users "can change an image's setting, style, or composition while retaining more of the details". Read that as a fix for the thing that made the previous generation unusable on real work: you'd ask for a different background and get a different product.

For anyone doing revisions against a brief, this is the whole ballgame. A client comes back wanting the same shot in evening light; if the model reinterprets the subject at the same time, you've lost the approved element and the round trip was wasted. Consistency across an edit beats a marginal gain in first-pass quality, because the second pass is where the day goes.

Inpainting works through the API's edits endpoint, so you can mask a region and aim the change at it. Read the guide before you trust the word inpainting, though: "Masking with GPT Image is entirely prompt-based. The model uses the mask as guidance, but may not follow its exact shape with complete precision." A mask is guidance, so budget for cleanup when the thing you need untouched is a logo or a face. 

Text and infographics inside the image

The system card claims improved accuracy and layout on infographics. Rendering legible, correctly spelled text inside a generated image has been the standing weakness of every diffusion-adjacent model, and it's the reason designers have kept generation out of anything that carries a word of copy.

Treat it as a direction of travel. OpenAI states the improvement, and nobody outside OpenAI has published a comparison yet. If you're weighing whether to put generated type in front of a client, run your own worst case, which is usually a long word at small size on a busy background.

Sketch, templates and comments, with a catch

Four additions land on the ChatGPT side. Sketch, invoked by typing @Sketch, lets you draw a rough visual reference instead of describing composition in words. Templates start you from a format such as a poster or merchandise, though they hadn't reached ChatGPT Work mode at launch. Comment-based edits let you place feedback on one part of an image and aim an edit at that spot. Shared prompts hand someone else the prompt behind an image to re-run with their own details.

Here's the catch. Those four are features of ChatGPT, not capabilities of the model. What the API exposes is the two models and their parameters: size, quality, output format and compression, background, moderation, and reference-image or mask inputs. None of the four appears anywhere in the image generation guide. So reach 2.5 through a third-party canvas and you get Flare and Sunburst without the interface built on top of them.

One distinction changes what you actually lose. Sketch-driven generation survives the trip: a sketch is just an image, and you can pass it to the API as a reference or a mask from any tool. The drawing surface is the ChatGPT-only part. If you already sketch in Procreate or on paper, you keep the technique and lose only the convenience of doing it in one window.

Resolution, rate limits and what it costs

Pricing is token-based and identical across both models: $5.00 per 1M text input tokens, $8.00 per 1M image input tokens, and $30.00 per 1M image output tokens, as of September 2026. Both model pages note text output isn't billed, since these models return images. The pricing page has the current rates.

What OpenAI does not publish is a per-image price. The guide's per-image dollar tables carry an explicit disclaimer that "the details below apply to earlier models, not Sunburst or Flare", leaving an interactive estimator in their place. Costing a 200-asset campaign takes a spreadsheet, and anyone quoting you a flat per-image figure for 2.5 got it somewhere other than OpenAI's price list.

Neither 2.5 model supports the Batch API. GPT Image 2 does, at $4.00 per 1M image input tokens and $15.00 per 1M image output, and it stays on the pricing page's batch tab while the 2.5 models are absent from it. If you generate in bulk on a schedule, moving up costs you that discount.

Recommended resolutions are 1024x1024, 1536x1024 and 1024x1536. Custom sizes must be multiples of 16, hold an aspect ratio between 1:3 and 3:1, and keep both edges under 3840 pixels. Above 2560x1440, OpenAI calls its own output experimental.

Rate limits bite before price does on a small account. The free API tier can't call these models at all. Tier 1 allows 5 images per minute, Tier 3 gets you to 50, and Tier 5 to 250. A freelancer on a fresh key generating variations will be watching a queue.

The grain nobody has fixed

The loudest complaint about 2.5 is texture, and it predates the release. Users have reported fine noise across skies, skin, grass and studio backgrounds since GPT Image 2.0. The comment picked up most widely put it as OpenAI making a fuss about 2.5 while leaving 2.0's biggest issue alone, and others in the same thread said the output looks about as noisy as before.

It survives outside screenshots. Eric Hal Schwartz tested four images for TechRadar on 10 September, prompting explicitly for smooth gradients and no visual noise, and reported "a persistent fine texture that crops up in skies, studio backgrounds, and low-light areas". He also declined to call it fatal: "It doesn't ruin the picture, but once I noticed it, I couldn't stop noticing it." One user, quoted in the same piece, said the artifacts make the model "mostly useless for production".

That test has limits. It's four images, first-person, in consumer ChatGPT, and it says nothing about Flare against Sunburst or about quality tiers, because nobody has published a matched test isolating the noise. OpenAI hasn't addressed the 2.5 complaints or shipped a fix, and a few users report cleaner output than before. The practical advice is unglamorous: generate a test frame with a large area of flat tone before you build a schedule around this model, especially if the work goes to print.

What the docs still get wrong

OpenAI's own documentation hasn't caught up. The models index now lists only Flare and Sunburst under image generation, dropping GPT Image 2, even though its page is live and the model isn't deprecated.

The deprecations page is the part to diary, and it's stranger still. gpt-image-1 shuts down on 2026-10-23. gpt-image-1.5, gpt-image-1-mini and chatgpt-image-latest all go on 2026-12-01. DALL-E 2 and DALL-E 3 stopped being accessible on 2026-05-12. If a plugin, a Shortcut or a studio script of yours calls any of those, it has weeks left.  

The odd part is what OpenAI recommends instead. Every deprecated model's replacement is listed as GPT Image 2, not 2.5, so the page hasn't been updated for its own new release. And the two DALL-E rows point at gpt-image-1 and gpt-image-1-mini, both of which are themselves on the shutdown list. Follow that migration advice literally and you'd port your work to a model that dies in October.

Where to reach GPT Image 2.5 outside the big platforms

Two days after launch, third-party access is thin. Most creative tools with an OpenAI integration still top out at GPT Image 2, because shipping a new model into a credit-priced product takes longer than announcing it. Four small tools already have 2.5 in the picker, and all meter by credits, so cost tracks quality tier and resolution. One warning covering three of the four: their pricing pages load with annual billing pre-selected, so the first number you see isn't the month-to-month price.

Morphic (morphic.com) is a canvas studio for image, video and audio work, with frame iteration and inbetweening aimed at motion and story people. Its credit rate card prices both 2.5 variants for text-to-image and image-to-image across all five quality tiers. Basic is $9/month for 1,100 credits and is the one plan priced the same monthly or annually; Standard is $29 monthly or $24 annually, Pro $49 or $45, Pro Max $185 or $170, and a $4 education tier includes Basic's model access with a school email, as of September 2026. The free plan is the trap: capped at 20 credits, 2.5 excluded, watermarked downloads, and terms granting output "solely for personal, non-commercial purposes", so you can't test this model on client work before paying. Paid output comes with a commercial licence that survives cancellation, though it's conditional: it covers only outputs whose fees were paid and kept, and a chargeback voids it. Browser only.

ComfyUI (comfy.org) has the best-documented 2.5 support here: its changelog added Flare and Sunburst to the OpenAI GPT Image node in v0.35.0 on 9 September, one day after launch, with four shipped workflow templates, 16 reference images and 1 to 8 images per request. It's also the only one of the four that asks you to build the pipeline yourself, which the templates soften but don't remove. The app is free under GPL-3.0; this node isn't. Partner nodes run on prepaid credits with no free usage, need a logged-in Comfy account, and don't accept user-supplied OpenAI keys. Two catches: Comfy hasn't published a credit rate for 2.5, so the node is billable at a price its own pricing table doesn't list, and the seed parameter is "Kept for API compatibility; the backend does not use it yet", so you can't reproduce a generation.

Melius (melius.com) wires each model into a node and passes its output downstream as prompt context. Its model docs list both Sunburst and Flare at 1K, 2K or 4K, editing from up to 16 connected reference images, which is how you hold a character or a product consistent across a set. Creator is $20/month, or $17 billed annually, as of September 2026, metered in credits; a free account gets model access but only trial credits. Its terms say you keep your outputs and can use them commercially on any plan, and that Melius doesn't train on your work, though it can't guarantee the data practices of the providers behind each model. Web and macOS 12 or later; Windows and Linux are waitlisted. Before you commit a job: there's no .psd, .ai or .eps import, and its own troubleshooting docs warn that a text node doesn't run when you create it, so a style reference you think is feeding an image node can be silently empty.

Lovart (lovart.ai) is a design agent that turns a brief into a set of on-brand assets on a canvas, pitched at campaign work, though it also ships single-asset editing tools. Unlike the others it puts GPT Image 2.5 on every tier including the free one, priced per generation from 3 credits for a 1K image, so the free plan's 30 daily credits are enough to try the model without a card. Paid tiers run $19/month on Starter or $16 billed annually, and $32 on Basic or $27 annually, as of September 2026, all with a commercial licence. Treat the top two tiers' headline rates with suspicion: Pro and Ultimate were discounted under a launch promotion when we checked on 10 September 2026, and both are first-year prices that step up on renewal. Web only.

What this changes at your desk

If you mostly edit, Sunburst is the reason to pay attention. Holding detail through a change of light or composition is the difference between a revision round that works and one you redo by hand anyway.

If you generate in volume, do the arithmetic before you move. Flare is faster and OpenAI rates it higher than GPT Image 2, but 2.5 has no Batch API, and losing a 50% batch discount buys back a lot of latency.

Whatever you pick, run the grain test first, on a frame with a large area of flat tone. The noise is a four-version-old complaint that this release didn't answer, and it shows up more at print size than on a phone. And if anything you rely on still calls gpt-image-1, put 23 October in the calendar now.