OpenAI pushed a new image model into ChatGPT on September 8, 2026, and the numbers behind it are bigger than the release itself. In the announcement post “Introducing ChatGPT Images 2.5,” the company said its image tools now handle more than 3 billion image generations a week across ChatGPT Images and the GPT-Image models available through the API. That volume, more than the population of Africa and Europe combined, is now running through a model OpenAI calls its “new state-of-the-art image model.”
The release lands alongside two new developer-facing models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst, and it rolls out immediately to every ChatGPT, ChatGPT Work, and Codex user on desktop, mobile, and web. There’s no waitlist and no tiered early-access period, a rollout pattern OpenAI has repeated for its last several consumer launches. Here’s what actually changed, what it costs, and how it stacks up against Google, Midjourney, and Adobe as the AI image generation market keeps compressing on price.
What Is ChatGPT Images 2.5?
ChatGPT Images 2.5 is OpenAI’s latest image generation and editing model built directly into the ChatGPT interface. It replaces ChatGPT Images 2.0 as the default engine behind every image request typed into the chat window, whether that’s a from-scratch generation, a photo edit, or a multi-turn creative session where a user refines the same image across several prompts.
According to OpenAI’s own announcement, the model brings faster generation, tighter prompt adherence, and better handling of text rendered inside images, an area where earlier generations of GPT-Image and rival tools alike have struggled. The company frames the release as a quality and speed upgrade rather than a new product category, but the scale it’s shipping at, 3 billion images a week, means even a modest quality bump touches an enormous number of users.
Coverage from outlets that tested the release early, including eesel.ai and Proje Defteri, describes three practical additions layered on top of the base model: a Sketch tool for drawing rough shapes that the model turns into finished images, reusable Templates for repeated image formats, and comment-based editing that lets a user annotate a specific region of an image and ask the model to change only that area. None of these three features appear in OpenAI’s own short-form announcement text, but both outlets independently describe testing them in the live product, which is the kind of independent corroboration that carries weight when a company’s own post is thin on detail.
Three Billion Images a Week: The Scale Behind the Launch
The 3-billion-a-week figure is the single most concrete number in OpenAI’s announcement, and it deserves more scrutiny than a passing mention. That volume spans two surfaces: the consumer ChatGPT Images feature and the GPT-Image family of models sold through the API to third-party developers. OpenAI did not break out what share comes from each side, so it’s not possible to say how much of that traffic is casual consumer generation versus production workloads from apps built on top of the API.
What the figure does confirm is that image generation has become a mainstream ChatGPT behavior, not a niche feature buried in a menu. For comparison, OpenAI has previously disclosed weekly active user counts for ChatGPT in the hundreds of millions. A 3-billion-image weekly run rate implies image generation is now a routine, repeat action for a large share of that base rather than a once-a-month novelty.
That scale also explains why OpenAI is emphasizing speed over splashy new capabilities in this release. A latency cut that shaves even a second off generation time compounds fast when it’s applied 3 billion times a week, both in server cost for OpenAI and in perceived responsiveness for users.
GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst: The Two New API Models
For developers, OpenAI is introducing two new models in the API rather than a single upgraded one. GPT-Image-2.5 Flare carries the same quality, editing, and speed improvements found in the consumer-facing ChatGPT Images 2.5 update. GPT-Image-2.5 Sunburst is positioned as the higher-precision option, offering an extra level of detail for demanding creative work at the cost of longer generation times.
That split mirrors a pattern OpenAI has used elsewhere in its model lineup: a fast, general-purpose default paired with a slower, higher-fidelity variant for workloads that can tolerate extra latency. OpenAI’s developer documentation lists both under the Images API, with gpt-image-2.5-sunburst as the identifier to use for standard image generation and editing calls. The Responses API can also route to a mainline model that supports the image generation tool, letting the tool handle model selection automatically rather than requiring the developer to hardcode a specific image model.
In practice, that means a developer building a product-photo generator that needs to run thousands of variations quickly would reach for Flare, while a design tool aimed at professional illustrators or marketing teams producing a handful of hero images a day would lean on Sunburst.
Sketch, Templates, and Comment-Based Editing: What Changed in the Interface
The consumer-facing changes matter as much as the model swap underneath them. Per reporting from eesel.ai and Proje Defteri, the Sketch tool lets a user draw a rough outline directly in the ChatGPT canvas, then hand that sketch to the model as a structural guide rather than a text prompt, a workflow closer to how a human art director would brief an illustrator than how most chatbot image tools currently work.
Templates give users a way to save a specific image format, aspect ratio, and style combination and reapply it to new prompts without re-describing every constraint each time. Comment-based editing lets a user click a region of an already-generated image and leave a note describing exactly what should change there, an approach closer to how designers mark up drafts in tools like Figma than how earlier chatbot-style image editing worked, where a full re-prompt was often needed to change one element.
Both outlets also note the model’s improved fidelity to reference photos and better handling of complex layouts with multiple text elements, addressing two of the most common complaints leveled at prior-generation image models across the industry, not just OpenAI’s own.
Pricing: What GPT-Image-2.5 Actually Costs Developers
OpenAI’s own announcement text does not spell out API pricing for Flare and Sunburst, but multiple independent developer-facing sites, including Proje Defteri, Kanaries, AIFreeAPI, eesel.ai, and Zenn.dev, all cite matching figures pulled from OpenAI’s official pricing documentation. Every one of those sources lands on the same numbers, and they all note something notable in itself: OpenAI did not raise prices for this generation. Flare and Sunburst bill at the identical rate card as the outgoing GPT-Image-2 model.
| Token Type | GPT-Image-2.5 Flare | GPT-Image-2.5 Sunburst | Prior Gen (GPT-Image-2) |
|---|---|---|---|
| Text input | $5.00 / 1M tokens | $5.00 / 1M tokens | $5.00 / 1M tokens |
| Cached text input | $1.25 / 1M tokens | $1.25 / 1M tokens | $1.25 / 1M tokens |
| Image input | $8.00 / 1M tokens | $8.00 / 1M tokens | $8.00 / 1M tokens |
| Cached image input | $2.00 / 1M tokens | $2.00 / 1M tokens | $2.00 / 1M tokens |
| Image output | $30.00 / 1M tokens | $30.00 / 1M tokens | $30.00 / 1M tokens |
Holding pricing flat while shipping a faster, more capable model is effectively a price cut on a per-quality basis, since developers get better output for the same token cost. It’s also a defensive move: Google, Midjourney, and Adobe have all been trimming per-image costs over the past year as the market for AI image generation gets more competitive, and a price increase from OpenAI right now would have handed rivals an easy talking point.
Rollout: Who Gets Access and When
Unlike some recent OpenAI launches that staged access by subscription tier, ChatGPT Images 2.5 went out to everyone at once. OpenAI’s community announcement states plainly that the update is “rolling out today to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web.” That includes free-tier ChatGPT users, not just paying subscribers, which is a meaningful departure from how OpenAI handled some of its more compute-intensive releases earlier in 2026.
The inclusion of Codex users is worth noting on its own. Codex is OpenAI’s coding-focused product line, and folding image generation access into it signals OpenAI expects developers to use image generation inside coding workflows, generating UI mockups, placeholder assets, or documentation graphics without leaving the tool they’re already writing code in.
How ChatGPT Images 2.5 Compares to Rivals
OpenAI isn’t launching into a quiet market. Every major AI lab has shipped a flagship image model refresh in the past seven months, and the competitive gap between them has narrowed to the point where feature parity, not raw output quality, is becoming the deciding factor for developers picking a provider.
Google: Nano Banana 2 (Gemini 3.1 Flash Image)
Google’s current flagship is Nano Banana 2, technically named Gemini 3.1 Flash Image, which launched February 26, 2026. Google has positioned it as the direct replacement for its older Imagen 4 line inside the Gemini API, and it’s now the recommended default across the Gemini app, Google Ads creative tools, and the Flow video platform. Google shipped Nano Banana 2 roughly six months before OpenAI’s ChatGPT Images 2.5, giving it a head start in developer mindshare, particularly among teams already standardized on the Gemini API for other tasks.
Midjourney: Version 8.2
Midjourney’s default generation model is currently V8.2, an incremental update on the V8 line the company shipped in 2025. Midjourney remains distinct from OpenAI and Google in that it operates primarily through Discord and its own web interface rather than a general-purpose developer API, which keeps it popular with illustrators and hobbyists but less commonly embedded directly into third-party software products.
Adobe: Firefly Image Model 5
Adobe’s newest flagship, Firefly Image Model 5, became generally available on March 19, 2026, bundled with expanded video and custom-model capabilities inside the Firefly app and Creative Cloud. Adobe’s pitch has always leaned on commercial safety, Firefly is trained on licensed and public-domain content specifically so enterprise customers can use its output without copyright exposure, a positioning OpenAI has not directly matched with equivalent public guarantees for ChatGPT Images.
| Provider | Current Flagship Model | Launch Date | Primary Access Point |
|---|---|---|---|
| OpenAI | ChatGPT Images 2.5 / GPT-Image-2.5 Flare & Sunburst | September 8, 2026 | ChatGPT app + API |
| Nano Banana 2 (Gemini 3.1 Flash Image) | February 26, 2026 | Gemini app + API | |
| Midjourney | V8.2 | Mid-2026 update | Discord + web app |
| Adobe | Firefly Image Model 5 | March 19, 2026 | Firefly app + Creative Cloud |
Public reactions from Google, Midjourney, or Adobe to OpenAI’s specific September 8 release hadn’t surfaced as of publication. That’s typical for this market: labs rarely comment directly on a rival’s launch, preferring to let their next release speak for them. The more telling signal will be whether Google or Adobe ship a countering update in the next six to eight weeks, roughly the cadence each company has kept to since early 2026.
Historical Context: From DALL-E to GPT-Image-2.5
OpenAI’s image generation lineage stretches back to the original DALL-E in 2021, followed by DALL-E 2 and DALL-E 3, which powered ChatGPT’s image features through much of 2023 and 2024. The shift to the GPT-Image naming convention marked a deliberate move to fold image generation into the same model family as OpenAI’s text and reasoning models rather than keeping it as a separate branded product. GPT-Image-2 followed, then ChatGPT Images 2.0, and now ChatGPT Images 2.5 with its Flare and Sunburst API variants.
That naming evolution reflects a broader trend across the industry: image generation is increasingly treated as one capability inside a general-purpose model rather than a standalone product line. Google folding Imagen into the Gemini family and deprecating it in favor of Nano Banana models follows the identical pattern. The days of a dedicated, separately-branded image model competing purely on aesthetic output are giving way to image generation as a feature embedded inside a broader multimodal assistant.
Market Impact: What This Means for the AI Image Generation Market
Flat pricing on a faster, more capable model puts pressure on smaller image-generation startups that have built businesses on being cheaper or faster than the big labs. If OpenAI, Google, and Adobe are all holding or cutting prices while improving quality, the margin for a niche provider to compete purely on cost keeps shrinking.
The 3-billion-image-a-week figure also has infrastructure implications. Image generation is far more compute-intensive per request than a typical text completion, and sustaining that volume at unchanged prices suggests OpenAI has made real efficiency gains in its serving infrastructure, not just its model architecture. That’s consistent with the 50% latency reduction OpenAI cites compared with Images 2.0. Faster generation per request directly cuts the compute cost of serving each image, which is likely what’s letting OpenAI hold the line on price.
For app developers building on top of GPT-Image-2.5, the two-model split (Flare and Sunburst) creates a genuine cost-versus-quality decision point for the first time in OpenAI’s image lineup, rather than a single one-size-fits-all model. Expect to see more products explicitly market “fast mode” and “detail mode” toggles as a direct result of this split, mirroring how video generation products already split fast preview renders from expensive final renders.
What the Announcement Says, Verbatim
OpenAI’s own language is worth reading closely rather than paraphrased, since the wording clarifies what the company is and isn’t claiming. On developer access, OpenAI stated: “For developers, we’re introducing two new models in the API,” according to OpenAI’s community announcement.
On what distinguishes the two new models, OpenAI said: “GPT-Image-2.5 Flare brings the same improvements in quality, editing, and speed, and GPT-Image-2.5 Sunburst offers an extra level of precision for detailed creative work with longer generation times,” per OpenAI’s community announcement.
On performance, OpenAI said: “We’ve also reduced image generation latency by up to 50% compared with Images 2.0, so you can generate images and refine your concepts more quickly,” according to OpenAI’s community announcement.
And on rollout, OpenAI confirmed: “ChatGPT Images 2.5 is rolling out today to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web,” per the same community announcement.
Notice what’s absent from OpenAI’s own wording: no mention of Sketch, Templates, or comment-based editing by name, and no published API pricing table in the announcement text itself. Both come from third-party testing and OpenAI’s separate developer pricing documentation, not from the launch post. That’s a minor but real gap between what OpenAI announced and what reviewers found when they actually opened the product, worth keeping in mind before treating every feature claim as officially confirmed by OpenAI.
Risks and Open Questions
A handful of questions remain unanswered by OpenAI’s announcement. The company has not published a breakdown of how the 3-billion-weekly-image figure splits between consumer ChatGPT usage and third-party API traffic, which makes it hard to gauge how much of that volume is commercial versus casual. OpenAI also hasn’t detailed content provenance measures, such as C2PA metadata tagging, for images generated by Flare and Sunburst, an area where Adobe has built much of its enterprise pitch and where regulators in the EU and elsewhere have been pushing for standardized disclosure on AI-generated media.
There’s also the open question of rate limits and abuse controls at this scale. Serving 3 billion images a week to a free-tier user base invites the same spam and misuse patterns OpenAI has had to police on the text side of ChatGPT, and image generation carries additional risk categories, deepfakes and impersonation chief among them, that OpenAI’s announcement doesn’t address in the material reviewed for this piece.
Predictions: Where AI Image Generation Goes From Here
- Google and Adobe will likely respond with their own updates within six to eight weeks, matching the release cadence both companies have kept through 2026 rather than letting ChatGPT Images 2.5 sit unanswered.
- Expect Flare-and-Sunburst-style two-tier splits (fast/cheap versus slow/precise) to become the default shape of image-model lineups industry-wide, following the same pattern already common in text and video generation.
- Flat API pricing despite a quality and speed upgrade suggests the next real competitive battle in AI image generation shifts to editing precision and multi-turn consistency rather than raw price-per-image, since the major providers are now converging on similar token rates.
- Provenance and watermarking disclosure will likely become a bigger public talking point over the next two quarters as regulatory pressure, particularly from the EU, pushes labs to publish clearer documentation on how generated images are tagged.
- Coding-tool integration, signaled here by Codex getting image access alongside ChatGPT, points toward image generation becoming a standard utility inside developer tools rather than a separate creative-app category, similar to how autocomplete became table stakes in code editors.
What Developers Should Do Next
Teams already building on GPT-Image-2 can switch to Flare or Sunburst without a pricing change, which makes this close to a drop-in upgrade for most existing integrations. The practical decision is which of the two new models fits a given workload: high-volume, latency-sensitive use cases (product thumbnails, bulk asset generation, real-time app features) fit Flare, while lower-volume, quality-critical use cases (marketing hero images, illustration work, print-ready assets) fit Sunburst’s longer generation time in exchange for added precision.
Teams evaluating providers from scratch now have four credible flagship options, OpenAI, Google, Midjourney, and Adobe, each with a model refreshed within the last seven months. That’s a meaningfully more competitive landscape than existed even a year ago, when OpenAI and Google were the only two labs shipping frequent, API-accessible image model updates.
Frequently Asked Questions
When did ChatGPT Images 2.5 launch?
OpenAI published its announcement, “Introducing ChatGPT Images 2.5,” on September 8, 2026, with the rollout beginning the same day.
What are GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst?
They are the two new image models OpenAI introduced in its API alongside the ChatGPT Images 2.5 update. Flare carries the same quality, editing, and speed improvements as the consumer release. Sunburst trades longer generation time for extra precision on detailed creative work.
How much does GPT-Image-2.5 cost through the API?
Based on OpenAI’s published API pricing and multiple independent developer write-ups, both Flare and Sunburst bill at $5.00 per 1M text input tokens, $1.25 per 1M cached text input tokens, $8.00 per 1M image input tokens, $2.00 per 1M cached image input tokens, and $30.00 per 1M image output tokens, the same rate card as the prior GPT-Image-2 model.
Who has access to ChatGPT Images 2.5?
OpenAI says the update is available to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web, including free-tier accounts, not just paid subscribers.
Is the Sketch tool an official OpenAI feature?
OpenAI’s own announcement text does not name a “Sketch” tool directly. Independent reporting from outlets including eesel.ai and Proje Defteri describes testing a sketch-to-image drawing tool, along with Templates and comment-based editing, inside the live ChatGPT Images 2.5 product.
How does ChatGPT Images 2.5 compare to Google’s Nano Banana 2?
Both are current-generation flagship image models from major labs. Google’s Nano Banana 2 (Gemini 3.1 Flash Image) launched February 26, 2026, roughly six months before ChatGPT Images 2.5. Direct head-to-head benchmark comparisons between the two had not been published as of this article’s publication.
Did OpenAI raise API prices with this release?
No. Multiple developer-facing sources confirm GPT-Image-2.5 Flare and Sunburst bill at the same per-token rates as the outgoing GPT-Image-2 model.
What happened to GPT-Image-2?
OpenAI’s announcement does not state that GPT-Image-2 is being deprecated or removed. It is presented as the prior generation, with Flare and Sunburst as the newer alternatives available at matching prices.
