OpenAI shipped two image models at one price — so the model name is not your cost decision, and the parameter that is swings 35x
TL;DR: On 8 September OpenAI shipped GPT Image 2.5 into ChatGPT for every tier and into the API as two models, gpt-image-2.5-flare and gpt-image-2.5-sunburst. The coverage led on sharper textures and up to 50% lower latency. Both are real. Neither is the number on your invoice. OpenAI held token pricing exactly flat against gpt-image-2 — $8 per million image input tokens, $30 per million image output — and gave both new models identical rates. So choosing Flare over Sunburst moves your rate card by zero. What moves it is the quality parameter, which now runs five rungs and spans $0.006 to $0.211 on the same 1024x1024 image. That is a 35x range on a knob that ships defaulted to auto. If you buy image generation by the thousand, this launch was not a price cut or a price rise. It was a widening of the range you can accidentally land in.
What actually shipped
Two API models, one snapshot date of 2026-09-08, and a consumer rollout.
| Model ID | Positioning | Latency | |
|---|---|---|---|
| Flare | gpt-image-2.5-flare | fast default, everyday generation | up to 50% lower than GPT Image 2 |
| Sunburst | gpt-image-2.5-sunburst | editing precision, premium creative | longer per generation |
OpenAI’s own developer guidance is a single sentence, and it is about workflow rather than quality tiering: choose Sunburst where editing precision matters most, and Flare for fast, high-quality everyday generation. Early customer Manus reported 2-4x speedups on Flare.
The capability additions are genuine. Supported sizes now reach 3840x2160 landscape and 2160x3840 portrait — about 8.3 megapixels. The quality ladder gained two rungs above the previous ceiling, xhigh and max. OpenAI cites sharper detail, richer textures, more natural lighting, better preservation of subjects from reference images, targeted edits across multiple turns without cumulative degradation, and transparent backgrounds in complex layouts.
On the consumer side, ChatGPT Images 2.5 went to all ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and web, alongside a Sketch tool, format templates, in-place image comments, and prompt sharing. That half of the launch is a feature drop, not a purchase decision — nobody has to buy anything.
The price that did not change, and why that is the story
Here is the line from OpenAI’s developer announcement that every buyer needs and almost no write-up quoted: the API carries the same token pricing as gpt-image-2.
| Token type | Rate per million |
|---|---|
| Image input | $8.00 |
| Image input (cached) | $2.00 |
| Image output | $30.00 |
| Text input | $5.00 |
Both Flare and Sunburst are on those rates. Identically.
This is unusual enough to sit with for a moment. A vendor shipping a two-model split almost always prices the split — a premium tier and a volume tier, with the rate card telling you which is which. OpenAI did the opposite. It differentiated on latency and behaviour and left the meter untouched, which is the same instinct visible in how it structured the Astra configuration surface a day earlier: the knob, not the SKU, is where the money moves.
The consequence for a buyer is precise and slightly counterintuitive. You cannot reason about Flare-versus-Sunburst cost from the price list, because the price list is the same document for both. But the two models do not cost the same. Sunburst runs longer generations; longer generations emit more output tokens; output tokens at $30 per million are what you are billed for. So Sunburst is more expensive than Flare for the same nominal image — the difference is simply invisible on the rate card and unmeasured in public. Neither OpenAI nor any independent source has published the per-image token delta. If you are moving volume to Sunburst, that measurement is yours to make before you commit, not after.
The 35x knob
The quality parameter is where the actual money is, and it now has five explicit rungs plus auto. Measured on a 1024x1024 image:
| Quality | Output tokens | Cost per image |
|---|---|---|
low | 196 | $0.006 |
medium | 439 | $0.013 |
high | 1,756 | $0.053 |
xhigh | 3,122 | $0.094 |
max | 7,024 | $0.211 |
From low to max is 35x — same model, same rates, same output dimensions. Set against that, the Flare/Sunburst decision is rounding error.
And the default is auto. OpenAI chooses on your behalf, and it has just been handed two new rungs above the old ceiling to choose from. Consider a pipeline generating 10,000 images a month. If auto settles on medium, that is $130. If it settles on high, $530. If a prompt pattern pushes it to xhigh, $940. Nothing in your code changed, no price announcement fired, and your bill quadrupled.
The operational advice is unglamorous and worth more than any benchmark in this launch:
- Pin
qualityexplicitly. Do not shipautoto production at volume. The default is a convenience for experimentation, not a cost control. - Test downward, not upward. Most teams assume they need
high. Generate a sample set atlowandmediumfirst and look at them. At $0.006,lowis cheap enough that being wrong about it costs nothing; at $0.211, being wrong aboutmaxon 10,000 images costs $2,110. - Watch the new size ceiling. 8.3-megapixel output at
maxis a materially different unit of consumption from a 1024x1024 atmedium, and the rate card will not warn you.
Where this leaves the competitive picture
The comparison everyone will reach for is Google’s Nano Banana Pro, and it does not resolve cleanly — because OpenAI’s ladder crosses Google’s flat rate rather than sitting above or below it.
| Cost per image | |
|---|---|
GPT Image 2.5 low | $0.006 |
GPT Image 2.5 medium | $0.013 |
GPT Image 2.5 high | $0.053 |
| Nano Banana Pro (batch/flex, 1K–2K) | $0.067 |
GPT Image 2.5 xhigh | $0.094 |
| Nano Banana Pro (batch/flex, 4K) | $0.12 |
| Nano Banana Pro (standard, 1K–2K) | $0.134 |
GPT Image 2.5 max | $0.211 |
| Nano Banana Pro (standard, 4K) | $0.24 |
Read down that table and the headline comparisons dissolve. At high, OpenAI undercuts Nano Banana Pro’s standard rate by roughly 2.5x and still comes in under its batch rate. At max, OpenAI is about 1.57x more expensive than Google’s standard rate. The crossover sits around xhigh. Any article telling you flatly that one is cheaper than the other has fixed the quality setting without telling you which one it fixed — and that choice, not the vendor, is doing the work.
This is the same shape this desk has flagged repeatedly through 2026: the rate card is increasingly not where vendors compete. Google’s Lyria 3.5 looked like a losing price until the API question was asked, and Astra’s per-task arithmetic only resolved once the context cliff was located. Image generation has now joined that pattern. The published per-token number is stable, honest, and nearly useless on its own.
For buyers weighing the wider field — Midjourney for artistic direction, Adobe Firefly for indemnified commercial work, FLUX for self-hosting — none of that calculus moves this week. What moved is the cost ceiling and floor available inside one vendor’s API. Our ranked guide to image generation tools and the Midjourney vs Nano Banana Pro comparison both carry the category framing; this launch changes the OpenAI row in it, not the shape of the table.
What we do not know yet
Two gaps are worth naming, because the launch coverage papered over both.
There are no independent benchmarks. As of 9 September, neither Flare nor Sunburst had third-party leaderboard scores. Everything about quality traces back to OpenAI’s own material. The best external signal is inherited: predecessor GPT Image 2 held the top text-to-image position on Artificial Analysis at Elo 1178 — a reasonable prior, not a measurement. Given that the rate card is unchanged, there is no cost to waiting for numbers before migrating production traffic. That restraint is the same one warranted when benchmark methodology itself gets rewritten mid-cycle.
The Flare/Sunburst token delta is undocumented. OpenAI told developers the rates are identical and told them Sunburst takes longer. It did not tell them what that costs. For a team choosing between the two on anything other than latency, that is the single number that matters, and it does not exist publicly yet.
The bottom line
OpenAI shipped a better image model and did not raise the price, which is straightforwardly good news. But the useful buyer takeaway inverts the headline. Because the rates are flat and identical across both models, the model name is the least consequential choice in this API. The quality parameter is a 35x cost multiplier shipping on a permissive default, and it just grew two new rungs at the expensive end.
If you run image generation at volume, the action item from 8 September is not a migration. It is opening your existing integration, finding where quality is set, and discovering whether you set it at all.
Cost figures are measured per 1024x1024 image at published token rates and will vary with prompt length, reference images, and output dimensions. Verify against your own usage before committing budget.
Frequently asked questions
Is GPT Image 2.5 more expensive than GPT Image 2?
On the rate card, no — the token rates are identical, and OpenAI said so explicitly in its developer announcement: the API carries 'the same token pricing as gpt-image-2'. Image input is $8 per million tokens with cached image input at $2, image output is $30 per million, and text input is $5 per million. What changed is the range of things you can ask for. GPT Image 2.5 added two quality rungs above the old ceiling, xhigh and max, and extended supported sizes up to 3840x2160 or 2160x3840 — about 8.3 megapixels. Both of those consume more output tokens, so a shop that upgrades and starts using the new headroom will pay more per image while paying the same per token. That is the distinction to hold onto: the price of a token did not move, but the number of tokens a default request can now consume did.
Should I use Flare or Sunburst?
Flare for volume, Sunburst for precision editing — and understand that the difference costs you money even though the rates are the same. Flare is OpenAI's fast default, quoted at up to 50% lower latency than GPT Image 2, with early customer Manus reporting 2-4x speedups. It is the right pick for creator and social content, visual search, prototyping, and anything high-volume. Sunburst deliberately trades speed for editing precision and is aimed at production campaign creative, product photography, and design work where a targeted edit must not degrade the rest of the frame. The trap is assuming that identical token rates mean identical bills. Sunburst runs longer generations, longer generations emit more tokens, and tokens are what you are billed for. Neither OpenAI nor any independent source has published a per-image token delta between the two, so if you are moving a high-volume pipeline to Sunburst, measure your own before you commit.
What does the quality parameter actually cost me?
More than any other decision in the API, and it is the one most teams leave on the default. Measured on a 1024x1024 image, low consumes 196 output tokens for about $0.006; medium 439 tokens for about $0.013; high 1,756 tokens for about $0.053; xhigh 3,122 tokens for about $0.094; and max 7,024 tokens for about $0.211. That is a 35x spread between the cheapest and most expensive rung on the same nominal image, from the same model, at the same token rate. The default is auto, which means OpenAI picks for you. On a pipeline generating tens of thousands of images a month, the difference between auto resolving to medium and auto resolving to high is the difference between a $130 bill and a $530 one. Pin the parameter explicitly, and test whether low or medium is sufficient for your use case before assuming it is not.
How does this compare to Google's Nano Banana Pro on price?
It depends entirely on which rung you run, which is the whole point. Nano Banana Pro charges about $0.134 per 1K or 2K standard image and about $0.24 at 4K, with Batch and Flex pricing roughly halving that to $0.067 and $0.12. Against those numbers, GPT Image 2.5 at high quality — $0.053 — undercuts Nano Banana Pro's standard rate by about 2.5x and still comes in below its batch rate. At low, $0.006 is roughly 11x cheaper than Nano Banana Pro batch. But at max, $0.211 is about 1.57x more expensive than Nano Banana Pro standard, though still under its 4K rate. So there is no single answer to which is cheaper: OpenAI's ladder crosses Google's at around the xhigh rung. If you are running a cost comparison between the two, comparing model to model without fixing the quality setting produces a meaningless result.
Are there independent benchmarks for GPT Image 2.5 yet?
No, and that is worth weighing before a migration. As of 9 September neither Flare nor Sunburst carried independent leaderboard scores. The available evidence is OpenAI's own launch material — sharper detail, richer textures, more natural lighting, better subject preservation from reference images, and multi-turn targeted edits that do not degrade the frame — plus its internal safety figures of 1.41% unsafe generation for Flare and 1.09% for Sunburst. The strongest external signal is inherited rather than measured: the predecessor, GPT Image 2, held the top position on Artificial Analysis text-to-image at an Elo of 1178. That is a reasonable prior for the successor being competitive, and it is not the same thing as a verified result. Given the rate card did not change, there is no financial penalty to waiting a couple of weeks for third-party numbers before moving production traffic.
Does the ChatGPT side of this launch change anything for non-developers?
It adds workflow features rather than a pricing decision. ChatGPT Images 2.5 rolled out on 8 September to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web, so there is no upgrade to buy and no new tier to evaluate. The additions are a Sketch tool that turns a rough drawing into a finished image, format templates for things like posters and merchandise, image comments for requesting specific changes in place, and prompt sharing. For a creator working by hand, the meaningful changes are the latency drop and the improved subject preservation across edit turns — the long-standing weakness where iterating on an image slowly destroyed the face you started with. None of the per-image cost arithmetic in this piece applies to you; it is an API concern only.
Sources
- OpenAI — Introducing ChatGPT Images 2.5 (announcement, 8 September 2026)
- OpenAI Developer Community — Introducing GPT Images 2.5 in the API and ChatGPT
- Simon Willison — Introducing ChatGPT Images 2.5 (8 September 2026)
- 9to5Mac — OpenAI releases ChatGPT Images 2.5 with 'sharper details' and 'more precise editing'
- CellCog — GPT Image 2.5 is here: Flare, Sunburst, what changed and what it costs
- OrcaRouter — GPT-Image-2.5 Flare vs Sunburst: new OpenAI image APIs
- OpenRouter — GPT Image 2.5 Sunburst API pricing and providers
- OpenRouter — Nano Banana Pro (Gemini 3 Pro Image) API pricing
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.