OpenAI rolled out ChatGPT Images 2.5 on September 8, 2026, alongside two new API models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst, according to OpenAI’s own release documentation, a count independently corroborated by outlets including Datanorth in their September 2026 coverage of the release. The launch landed in the middle of the busiest product week OpenAI has had all year, sandwiched between the September 3 debut of GPT-6 Astra and a wave of competing releases from Anthropic, Google, and Meta. Unlike Astra, which grabbed headlines for reportedly hitting a “Critical” cybersecurity threshold in OpenAI’s Preparedness Framework, the image update slipped out with comparatively little fanfare. That has since changed: developers have worked through the pricing sheets, creators have stress-tested the new editing tools, and OpenAI’s pricing page, in its October 14–17, 2026 listing, shows both new models at the same $8, $2, and $30 per 1 million token rates for image input, cached image input, and image output.
ChatGPT Images 2.5 is now live across every ChatGPT tier, including the free plan, plus ChatGPT Work and Codex, on desktop, mobile, and web, per OpenAI’s rollout notes. The update is not a new foundation model in the way GPT-6 Astra is. It is a targeted upgrade to how OpenAI generates and edits images, and it changes the economics of running image generation at scale through the API. Here is what actually shipped, what it costs, and how it stacks up against the rest of a crowded image-generation market.
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 Is ChatGPT Images 2.5? Launch Timeline and Availability
ChatGPT Images 2.5 is OpenAI’s latest image generation and editing engine, built into the ChatGPT interface and backed by two new API-facing models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. OpenAI describes the release as a state-of-the-art image model that delivers sharper detail, more precise editing, and generation speeds up to 50% faster than the prior Images 2.0 generation, according to OpenAI’s product documentation. On the API side specifically, OpenAI says GPT-Image-2.5 Flare carries that same up-to-50% latency reduction over GPT-Image-2, the model it directly replaces, a comparison OpenAI published alongside the broader Images 2.5 rollout in September 2026.
The rollout was immediate and broad. OpenAI pushed ChatGPT Images 2.5 to every consumer and business tier simultaneously rather than gating it behind a paid plan first, a departure from how some recent model launches have been staged. OpenAI underscored the scale behind that rollout: users now generate more than 3 billion images every week across ChatGPT Images and its GPT-Image API models, making it one of the highest-volume products the company ships. Codex, OpenAI’s coding-focused product, also picked up the new image capabilities, bringing the September 2026 rollout to three distinct product environments, ChatGPT, ChatGPT Work, and Codex, which matters for developers who use generated diagrams, mockups, or UI assets inside coding workflows.
The timing is notable. OpenAI’s own release tracker groups GPT-Image-2.5 alongside GPT-6 Astra, and the same first week of September 2026 saw Meta release Muse Spark 1.3 and Google roll out Gemini 3.8 Flash, both reported on September 2, alongside fresh model updates from Anthropic in the same window. Industry trackers counted between 9 and 22 new AI model releases across six or more providers in that single week, depending on how the count is scoped. ChatGPT Images 2.5 was one entry in what has become the densest cluster of frontier AI launches of 2026 so far.
Inside GPT-Image-2.5: What Flare and Sunburst Actually Do
OpenAI split the API side of the release into two named models: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Both sit on the same underlying architecture and share identical token pricing, including the $8 per million image input token rate that AI Tools Review verified in its September 2026 breakdown of the launch, according to OpenAI’s API pricing documentation. Flare is positioned as the general-purpose successor that carries forward the quality, editing, and speed improvements from ChatGPT Images 2.5 into the developer-facing endpoint. Sunburst is documented separately in OpenAI’s model reference, where it is described as an image-output model, meaning it returns generated images rather than text, billed under the same per-token structure as Flare. OpenAI’s own September 2026 documentation draws a further distinction between the two: Sunburst trades speed for fidelity, producing higher-precision output than Flare at the cost of longer generation times, which positions it as the pick for jobs where accuracy matters more than turnaround.
Neither model introduces a new resolution ceiling beyond what OpenAI’s image generation guide already documents for the 1024×1024 reference size used in its pricing tables. What changed is inference speed and how well the model preserves the identity of a subject across edits. OpenAI’s product notes claim the new engine edits only the specific region a user asks it to change, rather than regenerating the whole frame, which had been a recurring complaint about earlier GPT-Image and DALL-E-era models when users tried to make a single small correction to an otherwise-finished image.
Developers who already integrated GPT-Image-2 do not need to rewrite their pricing logic. The per-token rates are unchanged from the previous generation, according to OpenAI’s published rate card, and OpenAI is instead passing the improvement through as lower effective cost per image at common quality settings, plus new quality tiers that give developers more control over the cost-to-fidelity tradeoff than before.
Choosing Between Flare and Sunburst in the API
OpenAI states that GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare are available in the API. OpenAI’s model documentation describes GPT-Image-2.5 Flare as the default choice for most use cases. OpenAI also says Flare produces higher-quality images than GPT-Image-2 with 50% lower latency.
OpenAI says GPT-Image-2.5 Sunburst is built for premium visual workflows that benefit from tighter control across edits, and it points to creative and editing workflows like production-ready campaign creative or polished product imagery. The practical reading for a developer is to start with Flare and move specific jobs to Sunburst when edit control matters more than turnaround.
On cost, OpenAI’s pricing page, in its October 14–17, 2026 listing, shows both models at $8 per 1 million image-input tokens, $2 per 1 million cached image-input tokens, and $30 per 1 million image-output tokens. Text inputs are listed at $5 per 1 million tokens and $1.25 per 1 million cached tokens. OpenAI says image costs are token-based and vary with prompt tokens, input-image tokens, output dimensions, quality, and reference-image edits. Any per-image figure is therefore an estimate, so test your own prompts and reference images before committing to a volume budget.
One naming point is worth keeping straight when you write documentation or procurement notes. OpenAI’s announcement distinguishes the ChatGPT product name from the API model names, so ChatGPT Images 2.5 refers to the ChatGPT experience, while Flare and Sunburst are the models you call through the API.
ChatGPT Images 2.5 Pricing: Per-Token and Per-Image Costs
OpenAI’s token-based pricing for GPT-Image-2.5 Flare and Sunburst breaks down as $5 per million text input tokens, $1.25 per million cached text input tokens, $8 per million image input tokens, $2 per million cached image input tokens, and $30 per million image output tokens, according to OpenAI’s official pricing page, figures that both Datanorth and AI Tools Review independently confirmed in their own September 2026 pricing breakdowns of the two new models. Those rates carry over unchanged from GPT-Image-2, which means the sticker price on paper looks identical to the outgoing generation.
The practical difference shows up once those token rates are converted into a cost per generated image, with one caveat: OpenAI’s official rate card does not list a flat $0.006 per image, and says total image cost varies with token usage, dimensions, quality, and reference-image edits. For a standard 1024×1024 output, third-party pricing breakdowns estimate the cost at roughly $0.006 for a low-quality image, $0.013 for medium, $0.053 for high, $0.094 for a new “X-high” tier, and $0.211 for the new top “max” tier. The medium and high tiers land meaningfully below what the equivalent settings cost on GPT-Image-2, which ran roughly $0.011 for low, $0.053 for medium, and $0.211 for high at the same 1024×1024 size. In practice that means the medium tier dropped about 75% (from $0.053 to $0.013) and the high tier dropped about 75% (from $0.211 to $0.053), while “max” effectively becomes the new ceiling for teams that need maximum fidelity and are willing to pay for it.
| Quality Tier | Approx. Cost per 1024×1024 Image | Typical Use Case |
|---|---|---|
| Low | $0.006 | Draft thumbnails, rapid iteration |
| Medium | $0.013 | Standard app or content generation |
| High | $0.053 | Marketing assets, product mockups |
| X-High (new) | $0.094 | Client-facing creative work |
| Max (new) | $0.211 | Print-grade or hero imagery |
For teams running high-volume image pipelines, the cached image input rate of $2 per million tokens, a figure AI Tools Review’s September 2026 analysis singled out as one of the release’s most consequential line items, is the detail worth building around. Workflows that repeatedly reference the same source image (product photos being restyled dozens of times, for example) can cut input costs by 75% compared to the standard $8 per million rate simply by structuring requests to hit the cache. OpenAI’s pricing page, in its October 14–17, 2026 listing, shows the same $2 per 1 million cached image-input tokens alongside $8 for standard image input and $30 for image output, so early cost models built on launch-day numbers still hold.
New ChatGPT Features: Sketch, Templates, and Comment-Based Editing
The consumer-facing side of the release adds four interaction modes that did not exist in the prior ChatGPT image workflow. Templates let a user start from a preset structure through Images then Templates, pick a starting point, and have ChatGPT ask follow-up questions to fill in details, according to OpenAI’s release notes. Sketch, which OpenAI introduced in September 2026 and made available on mobile by typing @ in the message box and selecting the Sketch option, gives users a doodle-based workflow: someone draws a rough shape and describes what they want it to become, shifting part of the prompt from text into a spatial gesture.
The image viewer itself picked up inline editing and commenting, so a generated image can be opened full-screen and marked up with specific change requests rather than re-describing the entire scene in a new text prompt. OpenAI also added prompt sharing in September 2026, which lets a user publish the exact prompt behind an image so others can remix it into their own shared image variation. Taken together, these four features move ChatGPT’s image tools away from a pure prompt-in, image-out loop and toward something closer to a lightweight creative editor with AI generation underneath it. For a practical walkthrough of chaining these tools into a production pipeline, see our AI image editing workflow guide.
None of this is unique in concept. Competing tools have offered inpainting, region-based editing, and remixable prompts for a while. What is new is having all four bundled directly into the default ChatGPT interface that hundreds of millions of people already open weekly, rather than requiring a separate app or a paid image-specific product.
How ChatGPT Images 2.5 Compares to GPT-Image-2 and Images 2.0
OpenAI frames the jump from Images 2.0 to Images 2.5 as an efficiency and control release rather than a ground-up architecture change. The headline number is the claimed 50% reduction in generation latency, which matters most for interactive use cases like iterating on a design in real time rather than batch-generating assets overnight. OpenAI also says the new model does a better job holding onto the likeness of people and pets across edits, an area where GPT-Image-1 and DALL-E-era models were known to drift, subtly changing a face or a dog’s markings after a second or third edit pass.
Axios reporters who tested the update ahead of launch described the new engine as faster than its predecessor and better at preserving the likeness of people and pets, while also allowing more detailed editing, in hands-on impressions covering the rollout. PCMag’s own September 2026 testing reached a similar verdict, reporting that Images 2.5 preserves reference subjects more faithfully than its predecessor and cuts generation latency by roughly 50%. That lines up with OpenAI’s own framing: this is an iterative but real upgrade to speed and editing precision, not a wholesale replacement of the underlying image model family.
The localized-editing behavior is the change most likely to matter for day-to-day use. Earlier GPT-Image and DALL-E generations tended to regenerate large portions of an image even when a user asked for a small, specific change, like swapping a shirt color or removing a background object. OpenAI says Images 2.5 constrains edits to the requested region, which should reduce the number of retries needed to get a usable result and, by extension, reduce the effective cost per finished image even before accounting for the lower per-tier pricing.
Competitive Landscape: Where GPT-Image-2.5 Sits Against Rivals
OpenAI has not published head-to-head benchmarks against rival image models, and no independent lab has released a verified comparison at the time of writing. What is available is the pricing compiled in launch-week coverage across the major players, which gives a useful proxy for how OpenAI is positioning Images 2.5 in the market even without a shared quality benchmark.
| Model | Maker | Approx. Price per Image | Notable Strength |
|---|---|---|---|
| GPT-Image-2.5 (Flare/Sunburst) | OpenAI | $0.006 – $0.211 (tiered) | Localized editing, ChatGPT-native workflow |
| Nano Banana 2 | $0.067 | Deep Google Workspace integration | |
| Grok Imagine Image 2.0 | xAI | $0.04 | Fast turnaround, X platform tie-in |
| Seedream 5.0 Pro & Lite | ByteDance | $0.035 | Low-cost bulk generation |
| Midjourney V8.1 | Midjourney | Subscription-based | Stylized, artistic output |
Read against that table, OpenAI’s tiered structure is unusual. Rather than a single flat per-image rate, GPT-Image-2.5’s low and medium tiers undercut Nano Banana 2, Grok Imagine Image 2.0, and Seedream 5.0 on price, while the new X-high and max tiers sit above all three for teams that want the extra fidelity. That gives OpenAI a wider spread than competitors offering one price point, letting developers choose cheap-and-fast or expensive-and-precise from inside the same API rather than switching providers for higher-end work.
The strategic subtext is distribution. Nano Banana 2, Grok Imagine, and Seedream each rely on their parent platform’s existing audience, Google Workspace, X, and ByteDance’s app ecosystem respectively. OpenAI’s advantage is that ChatGPT Images 2.5 ships by default to every ChatGPT user, free tier included, which puts the new editing tools in front of a larger existing user base than any single rival can currently match through their own consumer surface.
Why OpenAI Shipped an Image Update During Its Biggest Launch Week
The image update landed five days after GPT-6 Astra, OpenAI’s flagship model for complex reasoning and coding, which launched September 3, 2026 at $10 per million input tokens and $50 per million output tokens, with a 1.05 million token context window and long-context requests billed at higher rates of $20 per million input and $75 per million output tokens. Astra also reportedly became the first OpenAI model to trigger the “Critical” cybersecurity threshold in the company’s Preparedness Framework, a classification that brought heightened access controls and scrutiny, and reportedly prompted OpenAI to pause further training while it evaluated the model’s capabilities. OpenAI has since released GPT-6.1 Sol on September 29, 2026, which it describes as a GPT-6-series model for coding, computer use, and professional work, priced at one-fifth of Astra’s standard API token prices.
Shipping ChatGPT Images 2.5 in the same window served a different purpose than Astra. Where Astra was aimed at enterprise reasoning and coding customers willing to pay a premium, Images 2.5 went straight to the free tier of ChatGPT, a product decision that reads as OpenAI reinforcing consumer engagement and everyday utility at the exact moment its flagship reasoning model was drawing scrutiny over safety classification. Bundling a widely-usable, low-controversy feature release alongside a higher-stakes model launch is a pattern OpenAI has repeated before, and it tends to keep the news cycle from concentrating entirely on the more sensitive release.
Safety Framing: The System Card and Deepfake Concerns
OpenAI published a system card for ChatGPT Images 2.5 through its Deployment Safety Hub, describing the release as a major step forward in image generation capabilities. System cards of this kind typically walk through disallowed content categories, impersonation and deepfake safeguards, and abuse-detection measures, following the format OpenAI has used for prior releases such as its GPT-5 system card, which lays out the Preparedness Framework methodology in detail.
Better likeness preservation is a double-edged improvement from a safety standpoint. The same capability that makes Images 2.5 useful for legitimate editing, keeping a person’s face consistent across multiple generated variations, is the capability that makes an image model more useful for creating convincing fake photos of real people. OpenAI has not disclosed a specific documented misuse incident tied to this release, and no credible report at the time of writing describes a controversy unique to Images 2.5. The company’s own materials frame the launch as balancing stronger editing tools against its existing safety guardrails, which is the same language it has used for prior generative media releases going back to the original DALL-E rollout.
Market Impact: What Changes for Developers and Businesses
For developers already building on the OpenAI API, the practical impact of ChatGPT Images 2.5 is a cost and quality upgrade that requires no migration work, since Flare and Sunburst use the existing token-based billing model. A simple API call for image generation under the new pricing looks like this:
curl https://api.openai.com/v1/images/generations \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-image-2.5-flare",
"prompt": "product photo of a ceramic mug on a wooden table, soft daylight",
"size": "1024x1024",
"quality": "high"
}'Teams running e-commerce catalogs, marketing asset pipelines, or in-app content generation stand to benefit most directly, since the medium and high tiers now undercut the equivalent GPT-Image-2 pricing while producing fewer failed edits that need to be discarded and regenerated. For businesses evaluating vendors, the addition of an X-high and max tier also means OpenAI can now compete for premium, print-grade or brand-critical image work that previously might have pushed a customer toward Midjourney or a specialized design tool instead.
On the consumer side, the free-tier rollout of Sketch, Templates, and comment-based editing lowers the skill floor for people who want to art-direct an image without writing a detailed text prompt. That matters commercially because prompt-writing has been one of the biggest friction points keeping casual users away from AI image tools, and OpenAI’s own model documentation now treats image generation as a first-class capability alongside text and reasoning rather than a bolt-on feature.
Historical Context: From DALL-E to GPT-Image-2.5
OpenAI’s image generation line has moved through several distinct generations since DALL-E’s 2021 debut. DALL-E 2 introduced more coherent, editable outputs in 2022, DALL-E 3 improved prompt adherence and integrated directly into ChatGPT in 2023 (the official DALL·E GPT was later scheduled for retirement on August 30, 2026, leaving ChatGPT Images as OpenAI’s recommended image generation and editing experience), and the GPT-Image family that followed shifted the underlying approach toward the same multimodal transformer architecture used in OpenAI’s language models, described in the same lineage of system cards OpenAI has published for models like GPT-4o. Each generation chipped away at a specific weakness: DALL-E 3 fixed prompt adherence, GPT-Image-1 improved text rendering inside images, and GPT-Image-2 pushed resolution and editing quality further.
GPT-Image-2.5 continues that pattern rather than breaking from it. The consistent theme across every generation has been narrowing the gap between what a user describes and what the model actually renders, plus reducing the number of attempts needed to get a usable result. Speed and localized editing are simply the next weaknesses in that chain, and OpenAI’s own materials position this release as evolutionary rather than a new foundation model, unlike the more architecturally significant jump the company is making with GPT-6 Astra on the text and reasoning side.
The Broader AI Image Generation Market in 2026
Image generation has become one of the more commoditized layers of the AI stack, with per-image pricing across major vendors now clustered within a few cents of each other at standard quality. That commoditization is exactly why OpenAI’s tiered pricing move matters: rather than compete purely on a flat per-image rate, which invites a race to the bottom, OpenAI is trying to differentiate through workflow, bundling generation with editing, templating, and sharing features inside a product people already use for other tasks.
Rivals are taking different paths to the same goal. Google’s Nano Banana 2 leans on Workspace integration, ByteDance’s Seedream 5.0 competes largely on raw cost for bulk generation, and Midjourney has stayed on a subscription model that trades API flexibility for a more curated, artistic aesthetic. None of the major players has converged on identical pricing or identical feature sets, which suggests the market is still deciding whether image generation will be sold as a metered API commodity, a subscription creative tool, or a free feature bundled into a larger AI assistant, the path OpenAI is now clearly betting on with the free-tier rollout of Images 2.5.
Predictions: What Comes Next for AI Image Generation
- Expect competitors to respond with their own tiered pricing structures within the next two quarters, moving away from flat per-image rates toward the low/medium/high/max model OpenAI just introduced.
- Video generation is the more likely next battleground for OpenAI’s consumer image tools, given the pattern of image features (Templates, Sketch) preceding equivalent video features in past release cycles.
- Enterprise customers running high-volume catalog or marketing pipelines will increasingly negotiate custom cached-token rates, since the $2 per million cached image input rate already signals OpenAI’s willingness to discount repeat-reference workloads.
- Regulatory attention on AI-generated imagery will likely intensify alongside, not because of, Images 2.5 specifically, as likeness-preservation improvements across the entire industry make deepfake detection a harder technical problem for platforms broadly.
- Expect OpenAI to eventually merge its image and reasoning model lines more tightly, following the same multimodal direction GPT-6 Astra represents on the text side, rather than maintaining separate Flare/Sunburst branding indefinitely.
What This Means for Creators and Small Businesses
For solo creators and small businesses, the practical takeaway is that a professional-grade image editing workflow, sketch-to-image, region-specific edits, and shareable prompt templates, now sits inside a free chat interface rather than a paid, specialized design tool. That is a meaningful shift for anyone who has been paying for a subscription-based image generator purely for its editing controls rather than its base generation quality.
The tradeoff is that free-tier access typically comes with usage limits, and OpenAI’s September 8, 2026 update left the existing image-generation limits unchanged, so businesses running anything beyond occasional generation will still land in the paid API tiers described above. The per-image cost at the low and medium tiers is cheap enough that a small business generating a few hundred product images a month would spend a few dollars, but that calculation changes quickly for anyone running the max tier at volume for print or brand-critical assets.
Frequently Asked Questions
When did ChatGPT Images 2.5 launch?
OpenAI rolled out ChatGPT Images 2.5 and the GPT-Image-2.5 Flare and Sunburst API models on September 8, 2026, to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web.
What is the difference between GPT-Image-2.5 Flare and Sunburst?
Both are API-facing models built on the same underlying engine and share identical token pricing. Flare carries forward the general quality, editing, and speed improvements from ChatGPT Images 2.5, while Sunburst is documented as OpenAI’s image-output model for the same API tier.
How much does ChatGPT Images 2.5 cost per image?
OpenAI bills by tokens rather than a flat per-image price, so these are estimates: a standard 1024×1024 image costs approximately $0.006 at low quality, $0.013 at medium, $0.053 at high, $0.094 at the new X-high tier, and $0.211 at the new max tier.
Is ChatGPT Images 2.5 free to use?
Yes, it is available on ChatGPT’s free tier as well as paid plans and ChatGPT Work, according to OpenAI’s rollout notes, and the September 8, 2026 update left existing image-generation limits unchanged.
How is ChatGPT Images 2.5 different from GPT-Image-2?
OpenAI says Images 2.5 generates images up to 50% faster, better preserves the likeness of people and pets across edits, and edits only the specific region requested rather than regenerating the whole image. Token pricing is unchanged, but effective per-image cost at common quality tiers is lower.
What new features did ChatGPT get alongside Images 2.5?
OpenAI added Templates for starting from a preset structure, Sketch for drawing rough shapes on mobile, inline editing and commenting in the image viewer, and prompt sharing so others can remix a published prompt.
How does ChatGPT Images 2.5 compare to Google Nano Banana 2 or Midjourney?
OpenAI has not published direct benchmarks against competitors. On price, GPT-Image-2.5’s low and medium tiers undercut Nano Banana 2 ($0.067 per image), Grok Imagine Image 2.0 ($0.04), and Seedream 5.0 ($0.035), while its new X-high and max tiers target higher-fidelity work above those price points.
Why did OpenAI launch an image model the same week as GPT-6 Astra?
OpenAI has not stated a specific reason. Astra targeted enterprise reasoning and coding customers and drew scrutiny for reportedly triggering OpenAI’s “Critical” cybersecurity threshold, while Images 2.5 shipped broadly to the free consumer tier, spreading the week’s news across a higher-stakes release and a lower-controversy, widely usable feature update.
