OpenAI pushed out a new image generation stack on September 8, 2026, and the timing tells its own story. Less than five months after Business Insider reported the launch of ChatGPT Images 2.0 with “thinking capabilities,” the company has already moved to version 2.5, adding a drawing tool called Sketch and splitting its flagship image API into two distinct models. The release, confirmed by OpenAI’s own developer changelog and covered by outlets including The Verge, marks one of the fastest iteration cycles OpenAI has run on any single product line this year.
For a company whose core image model had barely settled into GPT Image 2, shipping a follow-up with two specialized siblings, gpt-image-2.5-flare and gpt-image-2.5-sunburst, at the same token price signals something specific: OpenAI is treating image generation as a battleground it can’t afford to slow down on, not a feature it can leave on autopilot while text models grab headlines.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What OpenAI Actually Announced on September 8
The rollout has three separate pieces, and it’s worth separating them because they land on different products. First, ChatGPT Images 2.5 is the new consumer-facing generation engine inside the ChatGPT app itself. Second, Sketch is a new interaction mode bundled into that release, letting users draw directly in the chat window as a visual reference rather than describing everything in text. Third, and separately, OpenAI shipped two new models into its developer-facing Image API and Responses API: GPT Image 2.5 Flare and GPT Image 2.5 Sunburst.
According to OpenAI’s own developer changelog, the company “released GPT Image 2.5 Sunburst and GPT Image 2.5 Flare for image generation and editing through the Image API and the Responses API image generation tool.” That single sentence is doing a lot of work: it confirms this isn’t a UI skin on top of the old model, but a genuine new pair of models sitting inside the API surface developers already build against.
On the consumer side, OpenAI’s community announcement states that ChatGPT Images 2.5 is rolling out today to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web, putting the new engine in front of three distinct product tiers on three separate interfaces simultaneously as of the September 2026 launch. OpenAI’s own framing of that September 2026 rollout is explicit that it covers all tiers at once rather than staggering access to paying power users first, which is consistent with how OpenAI has handled most of its recent ChatGPT feature drops this year.
How the Sketch Feature Actually Works
Sketch solves a problem anyone who has fought with an image generation prompt already knows: text is a clumsy way to describe spatial layout. OpenAI describes the feature directly: “Sketch is a new feature that lets you draw directly in ChatGPT as a reference for your final image.” When OpenAI added the Sketch tool in September 2026, it built the canvas specifically for touchscreen image references, so a finger-drawn sketch on a phone carries the same compositional weight as a mouse-drawn one on desktop. The pitch is that a rough drawing carries information a paragraph of text can’t easily convey, like where a couch sits relative to a window, or the exact silhouette of a jacket collar.
To activate it, users type “@Sketch” into the ChatGPT message box, which opens a drawing canvas. On mobile, OpenAI’s help documentation says to type “@” in the message box and select Sketch, then draw the idea and describe the image afterward. Once the sketch is submitted alongside a text description, the model uses the drawing as compositional guidance while filling in detail, color, and style from the prompt.
OpenAI’s framing of the use case is telling: “Now you can quickly sketch the layout of a room, the contour of an outfit you’re concepting, or just a funny doodle, and ChatGPT will turn your rough art into a complete image.” That covers three distinct audiences in one sentence, interior design mockups, fashion concepting, and casual doodling, which suggests OpenAI is positioning Sketch as a mainstream feature rather than a niche professional tool.
The Verge’s coverage frames Sketch in blunter consumer terms, describing it as a way to turn rough drawings into finished AI images without needing to describe every visual detail in words. That’s a meaningfully different pitch from prompt engineering: instead of learning to write better prompts, users draw what they mean.
Meet GPT Image 2.5 Flare and Sunburst
The API side of the release is arguably more consequential for developers than Sketch is for casual users, because it changes how builders choose an image model for production workloads. Instead of one flagship model handling every job, OpenAI now offers two models split by workload type.
OpenAI’s own guidance in the developer changelog draws the line clearly: “Use Sunburst for workflows where editing precision matters most, or Flare for fast, high-quality everyday image generation.” That’s a direct instruction to developers on which model to pick, and it mirrors a pattern OpenAI has used elsewhere in its lineup, offering a fast default and a slower, higher-fidelity option for cases that need it.
Developer analysis site OrcaRouter, which published a detailed breakdown the same day, describes the split this way: “Flare is positioned as the fast, default choice for everyday generation,” while Sunburst trades speed for control on detailed editing tasks. OrcaRouter also reports that OpenAI’s internal evaluations indicate Flare generates roughly 2-4x faster than GPT Image 2 in practice, and OpenAI’s own September 2026 messaging puts a public figure on that claim: up to 50% lower latency than ChatGPT Images 2.0, the generation Flare and Sunburst are directly succeeding.
Both models reportedly improve on two specific weaknesses developers have complained about in prior OpenAI image models: subject preservation across edits (keeping a person’s face or a product’s exact shape consistent when a reference photo is supplied) and multi-turn editing stability, where earlier changes to an image stay locked in place instead of drifting as new edit instructions come in – a gain OpenAI explicitly credited to the September 2026 release for improving multi-turn editing consistency. OpenAI’s own September 2026 framing distills the upgrade into a wider set of headline improvements than past cycles: sharper details, richer textures, and more natural lighting, layered on top of more precise editing and faster generation. Both Sunburst and Flare also add new “xhigh” and “max” quality settings in the API, giving developers more control over the speed-versus-fidelity tradeoff on a per-request basis.
Pricing: Same Token Rate, More Capability
The pricing story here is unusual, and it’s the detail most likely to matter to engineering teams deciding whether to migrate. According to OrcaRouter’s analysis, both new models bill on the exact same token-based rate card OpenAI used for GPT Image 2: image tokens at $8.00 per million input tokens and $30.00 per million output tokens, with text prompt tokens billed separately at $5 per million input and $10 per million output. Two other trackers corroborate those figures independently: Datanorth.ai confirmed the $30-per-million-output-token rate applies to both 2.5 API models in its September 2026 pricing breakdown, and developer documentation site Apidog separately verified the $8-per-million image-input and $5-per-million text-input rates that same month.
OrcaRouter’s headline framing captures why this matters: it calls GPT Image 2.5 “the rare image-model upgrade that costs nothing on paper,” since developers get claimed improvements in latency, editing precision, and detail sharpness without OpenAI raising the per-token price. That said, OrcaRouter also flags an important caveat: OpenAI has not published exact tokens-per-image figures for Flare, so the real-world per-image cost compared to GPT Image 2 is not fully verifiable from the outside. Teams running high volume through the API will want to benchmark their own typical prompt-and-output size before assuming costs are identical in practice.
GPT Image 2.5 Flare vs. Sunburst vs. GPT Image 2
Here’s how the three models stack up on the details OpenAI and independent analysis have confirmed so far:
| Model | Best for | Speed vs. GPT Image 2 | Image token price (input/output) | Access |
|---|---|---|---|---|
| GPT Image 2 (previous gen) | General-purpose generation and editing | Baseline | $8.00 / $30.00 per million | Image API, Responses API |
| GPT Image 2.5 Flare | Fast, high-quality everyday generation | Up to 50% lower latency (OpenAI); ~2-4x in OrcaRouter tests | $8.00 / $30.00 per million | Image API, Responses API |
| GPT Image 2.5 Sunburst | Precision editing, detailed creative work | Slower than Flare, prioritizes accuracy | $8.00 / $30.00 per million | Image API, Responses API |
The takeaway for developers currently on GPT Image 2 is straightforward: there is no pricing penalty for testing the new models, so the migration decision comes down to whether Flare’s speed or Sunburst’s editing precision actually improves output quality for a given workload, not whether the switch is affordable.
Where ChatGPT Images 2.5 Is Available Right Now
Availability is broader than a typical OpenAI feature drop, spanning both the consumer chat product and the developer platform on the same day.
| Surface | What’s new there | Confirmed rollout |
|---|---|---|
| ChatGPT (app, web, desktop, mobile) | ChatGPT Images 2.5 engine, Sketch, new templates, on-image comments, prompt sharing | Live for ChatGPT, ChatGPT Work, and Codex users |
| Image API | GPT Image 2.5 Flare and Sunburst for generation and editing | Live, same token pricing as GPT Image 2 |
| Responses API | Image generation tool updated to use Flare/Sunburst | Live per OpenAI’s developer changelog |
OpenAI’s release notes also mention additional consumer-side extras shipping alongside Sketch: layout templates for common formats like flyers and product photography, the ability to leave comments directly on a generated image, and a way to share prompts with other users. None of these are positioned as paid add-ons in the material OpenAI has published. They appear to be part of the same ChatGPT Images 2.5 rollout rather than a separate tier.
In OpenAI’s Own Words
OpenAI has not put a named executive in front of this release the way it sometimes does for larger model launches. The public-facing language instead comes from the company’s own blog post and its developer community announcement. Five statements from those two sources capture what OpenAI is telling users and developers directly:
“We’re introducing Sketch, a new feature which lets you draw right in ChatGPT and use it as a visual guide.”
OpenAI, official announcement, Introducing ChatGPT Images 2.5
“Now you can quickly sketch the layout of a room, the contour of an outfit you’re concepting, or just a funny doodle, and ChatGPT will turn your rough art into a complete image.”
OpenAI, official announcement, Introducing ChatGPT Images 2.5
“Use Sketch to draw right in ChatGPT and show it exactly what you have in mind.”
OpenAI Developer Community, Introducing GPT Images 2.5 in the API and ChatGPT
“ChatGPT Images 2.5 is rolling out today to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web.”
OpenAI Developer Community, Introducing GPT Images 2.5 in the API and ChatGPT
“We’re also introducing two new models in the API: GPT-Image-2.5 Flare brings the same quality, editing, and speed improvements, while GPT-Image-2.5 Sunburst adds precision for detailed creative work, with longer generation times.”
OpenAI Developer Community, Introducing GPT Images 2.5 in the API and ChatGPT
Notice what’s absent: no named spokesperson, no CEO soundbite, no framing around AGI or existential stakes. Compared to how OpenAI has talked about some of its language model releases this year, the tone here is deliberately product-manager plain. That itself is a signal about how OpenAI is positioning image generation, as a maturing, iterative product line rather than a headline-grabbing frontier announcement.
Historical Context: A Fast-Moving Image Model Lineage
To understand why this release matters, it helps to see how quickly OpenAI has been iterating on image generation compared to its earlier cadence. Business Insider’s reporting from April 2026 covered ChatGPT Images 2.0, describing a model OpenAI said had “thinking capabilities,” including the ability to crawl the internet and double-check its own output before returning an image. That framing was about trustworthiness, addressing concerns that AI-generated images could pass as real photographs. OpenAI’s launch of that April 2026 generation turned out to be short-lived at the top of the lineup: by the September 2026 Images 2.5 release, 2.0 had already been superseded as the newest generation.
Five months later, the conversation has shifted from “can we trust what this generates” to “how fast and precise is the generation itself.” That’s a meaningful pivot: OpenAI spent its 2.0 cycle on credibility and verification framing, and its 2.5 cycle on raw performance and editing control. SiliconANGLE’s own coverage of the September 2026 launch frames it plainly as a generational handoff, reporting that Images 2.5 replaces the April 2026 Images 2.0 release outright rather than running alongside it. It suggests the credibility problem hasn’t gone away, but OpenAI judged that speed and precision were the more urgent gaps holding back adoption among developers building production image tools.
| Release | Date | Headline focus | Source |
|---|---|---|---|
| ChatGPT Images 2.0 | April 2026 | “Thinking capabilities,” self-verification of generated images | Business Insider |
| GPT Image 2 | Earlier 2026 (predecessor generation) | Single flagship image model for API generation and editing | OpenAI developer changelog |
| ChatGPT Images 2.5 + GPT Image 2.5 Flare/Sunburst | September 8, 2026 | Sketch drawing input; split fast/precision API models; same pricing | OpenAI, The Verge, OrcaRouter |
The pattern across this timeline is acceleration, not just in model quality but in release frequency. A company that once shipped a major image model update roughly once or twice a year is now iterating on a roughly quarterly basis, at minimum for its image generation line.
What This Means for the Broader Image Generation Market
OpenAI isn’t operating in a vacuum here. Seeking Alpha’s coverage of the release tagged it as relevant to a specific set of public tickers, including Adobe and Alphabet, signaling that investors read this as a competitive move against Adobe’s Firefly tools and Google’s image generation stack rather than a standalone consumer feature. The financial press treating a ChatGPT feature update as market-relevant to two other public companies stands out on its own. It means Wall Street now watches OpenAI’s image generation cadence the way it watches chip announcements or cloud earnings.
The competitive pressure runs in multiple directions at once. Google has been pushing its own image generation tools aggressively this year, and Adobe has continued layering generative features into Firefly for its existing Creative Cloud base. Separately, xAI’s Grok Imagine and various open and semi-open image models have been competing largely on price per image rather than on editing precision. OpenAI’s decision to hold pricing flat while improving speed and editing quality reads as a direct response to that price pressure: rather than cutting prices to compete on cost, OpenAI is trying to make the existing price look more generous by increasing what a developer gets for it.
Midjourney remains a wildcard in this comparison because it operates on a different business model, subscription access rather than API-metered tokens, which makes direct pricing comparisons to OpenAI’s per-token structure difficult without oversimplifying. What’s clear is that the image generation market has split into roughly two lanes: consumer-subscription tools optimized for aesthetic output, and API-first platforms like OpenAI’s, competing on latency, editing control, and integration into existing developer workflows. GPT Image 2.5 is squarely a move to strengthen OpenAI’s position in the second lane while using Sketch to make a play at the first.
Developer Reaction and Migration Considerations
For teams already integrated with GPT Image 2, the migration calculus is unusually low-friction on paper. Because Flare and Sunburst share the same token pricing structure as GPT Image 2, there’s no cost barrier to testing the new models against existing workloads. OrcaRouter’s guidance to developers is to default new everyday generation and editing calls to Flare, reserving Sunburst specifically for workflows where subject fidelity and edit precision outweigh speed, such as product photography retouching or design mockups that go through multiple rounds of revision.
The unresolved question for production teams is the tokens-per-image ratio. Because OpenAI has not published exact figures for how Flare’s and Sunburst’s token consumption compares to GPT Image 2 on identical prompts, engineering teams running meaningful volume through the API should benchmark actual cost per generated image on their own representative workload before assuming the “same price” framing holds exactly at scale. A model that generates in fewer tokens but takes more attempts to nail a specific edit could still end up costing more or less than advertised, depending on use case.
Code Reference: Calling GPT Image 2.5 Flare via the Image API
Based on OpenAI’s developer changelog, the new models slot into the existing Image API and Responses API image generation tool using the same call pattern developers already use for GPT Image 2, just swapping the model identifier:
from openai import OpenAI
client = OpenAI()
response = client.images.generate(
model="gpt-image-2.5-flare",
prompt="A minimalist product photo of a ceramic mug on a linen background",
quality="high",
size="1024x1024"
)
image_bytes = response.data[0].b64_jsonSwitching to Sunburst for a precision editing task is a matter of changing the model string:
response = client.images.edit(
model="gpt-image-2.5-sunburst",
image=open("product_reference.png", "rb"),
prompt="Keep the exact product shape and logo placement, change the background to a marble countertop",
quality="xhigh"
)Because both models bill on the same rate card as GPT Image 2, existing cost-tracking and rate-limit logic built around the previous model generally carries over without rearchitecting, at least until actual per-image token consumption on the new models is independently benchmarked at scale.
Safety and Trust Questions Still Loom
None of the material OpenAI has published specifically about ChatGPT Images 2.5, Sketch, Flare, or Sunburst raises new safety or misuse concerns in the announcement itself. That’s worth flagging precisely because OpenAI’s April 2026 messaging around ChatGPT Images 2.0 leaned heavily on “thinking capabilities” designed to help the model double-check its own outputs and reduce the risk of convincing fake photographs, per Business Insider’s reporting. The 2.5 announcement doesn’t repeat that framing or explicitly claim additional safety improvements, which leaves an open question about whether the verification capabilities introduced in 2.0 carry forward unchanged into Flare and Sunburst, or whether they were deprioritized in favor of speed and editing precision. OpenAI’s public materials for this release don’t address that directly, so it remains a gap rather than a confirmed fact in either direction.
Predictions: Where OpenAI’s Image Strategy Goes From Here
Based on the pattern visible across ChatGPT Images 2.0 and 2.5, and the broader competitive pressure from Google and Adobe, a few things look likely over the next two to three quarters:
- More workload-specific model splits. Having proven out the Flare/Sunburst fast-versus-precise split for images, OpenAI is likely to apply the same pattern elsewhere in its generative media stack rather than maintaining single do-everything models.
- Sketch-style input methods spreading to other modalities. If Sketch performs well as a way to bypass clumsy text prompting, expect similar direct-manipulation input methods (voice annotation, gesture-based editing) to show up in future ChatGPT updates.
- Pricing pressure holds steady rather than dropping. Given that OpenAI chose to add capability at flat pricing rather than cut prices outright, a direct price war on image generation looks less likely in the near term than a features-for-the-same-price arms race with Google and Adobe.
- Faster release cadence becomes the norm. Five months between Images 2.0 and 2.5 suggests OpenAI now treats image generation improvements as an ongoing series rather than annual flagship drops, meaning a 2.6 or 3.0-class update before the end of 2026 would not be surprising.
- Scrutiny over verification claims will resurface. Because the 2.5 release is quieter on the “thinking” and self-verification messaging that defined 2.0, expect journalists and researchers to specifically ask OpenAI whether that safeguard still applies to Flare and Sunburst outputs.
These are informed projections based on the release pattern and competitive dynamics described above, not confirmed OpenAI roadmap items.
The Bigger Picture for ChatGPT Users and Developers
Step back from the model names and token prices, and the ChatGPT Images 2.5 release says something about where OpenAI thinks the real friction in AI image generation still lives. It’s not primarily in output quality anymore, most modern image models can produce convincing results from a good prompt. The friction is in translation: turning what’s in someone’s head into a prompt precise enough to get the result they actually wanted, and then editing that result without the model forgetting what it just did.
Sketch attacks the first problem by letting users skip language entirely for spatial ideas. Sunburst’s precision-editing focus attacks the second by making multi-turn edits behave more predictably. Flare, meanwhile, is a bet that raw speed matters more than incremental quality gains for the bulk of everyday generation requests, which is a reasonable assumption for a company watching competitors compete increasingly on cost and turnaround time rather than pure image fidelity.
Whether this specific combination of features moves market share away from Adobe, Google, or Midjourney is impossible to say from a same-day announcement. What is clear is that OpenAI shipped a genuinely substantive update, two new production models plus a new consumer interaction pattern, on a five-month cycle, at unchanged pricing. For a market that had started to treat image generation as a solved, commoditized feature, that’s a reminder the competition is still very much active.
Frequently Asked Questions
What is ChatGPT Images 2.5?
ChatGPT Images 2.5 is OpenAI’s updated image generation experience inside the ChatGPT app, launched September 8, 2026. It introduces a drawing tool called Sketch and is powered on the backend by the same model family as the new GPT Image 2.5 Flare and Sunburst API models.
What does the Sketch feature do?
Sketch lets users draw directly inside the ChatGPT chat window as a visual reference for the image they want generated. Activated by typing “@Sketch” (or “@” and selecting Sketch on mobile), it lets ChatGPT combine a rough drawing with a text description to produce a finished image.
What are GPT Image 2.5 Flare and Sunburst?
They are two new image models available through OpenAI’s Image API and Responses API. Flare is optimized for fast, everyday image generation. Sunburst is optimized for precision editing tasks where accuracy matters more than speed, and takes longer to generate.
How much do GPT Image 2.5 Flare and Sunburst cost?
Both models bill on the same token rate as GPT Image 2: $8.00 per million image input tokens and $30.00 per million image output tokens, plus $5 per million input and $10 per million output for text prompt tokens, according to analysis published by OrcaRouter. OpenAI has not published exact tokens-per-image figures, so real-world per-image cost may vary by workload.
Is ChatGPT Images 2.5 faster than GPT Image 2?
OpenAI claims up to 50% lower latency for the new generation compared to GPT Image 2. Developer analysis from OrcaRouter reports internal evaluations showing Flare generating roughly 2 to 4 times faster than GPT Image 2 in practice, though results can vary by prompt complexity and requested quality setting.
Who can use ChatGPT Images 2.5 right now?
According to OpenAI’s own community announcement, the rollout covers ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web as of the September 8, 2026 launch. Developers can access Flare and Sunburst through the Image API and Responses API immediately.
How does this compare to ChatGPT Images 2.0?
ChatGPT Images 2.0, covered by Business Insider in April 2026, focused on giving the model “thinking capabilities” to double-check its own output and reduce convincing fakes. ChatGPT Images 2.5 shifts the focus to speed, editing precision, and a new sketch-based input method, without OpenAI’s public materials explicitly repeating the same self-verification claims made for 2.0.
Does this affect competitors like Google, Adobe, or Midjourney?
Seeking Alpha’s coverage of the release tagged it as relevant to Adobe and Alphabet’s public stock tickers, signaling investor interest in how it affects Firefly and Google’s image tools. No official response from Google, Adobe, or Midjourney has been documented in connection with this specific release as of publication.
