OpenAI shipped GPT Image 2.5 on September 8 with two models, Flare and Sunburst, and published the same token price for both. That detail is easy to skim past, and it is the whole story. When the fast model and the slow model cost the same per token, the question stops being “which one is better” and becomes “which job am I doing, and what does a usable image actually cost me.” That is exactly the shift the text models went through two years ago, and it is now finished on the image side too.
If you generate images as part of a product, a content pipeline, or internal work, the 2026 field has settled into four tiers that map to four different jobs. Picking “the best image generator” is the wrong exercise. Here is the field, what each tier is actually for, and the numbers that decide it.
What GPT Image 2.5 actually changed
The OpenAI API lists two new image models: gpt-image-2.5-flare and gpt-image-2.5-sunburst. Flare is the fast default, delivering higher quality than the prior GPT Image 2 at roughly 50% lower latency than the 2.0 generation. Sunburst spends more compute per image for extra fidelity on intricate detail and lower drift on edits. Both are reachable through ChatGPT, ChatGPT Work, Codex, and the API.
Here is the part worth sitting with. Both models bill at the same published rates: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens, with cached input discounted. Price is not the lever between Flare and Sunburst. Flare is the sane default for creator content, product imagery, visual search, and high volume work. Sunburst is a specialist you reach for only when you have measured that its lower drift matters on a specific edit task, per the WaveSpeed breakdown of the two endpoints.
Two things to know before you budget around it. First, GPT Image 2.5 exposes the same effort dial the frontier text models use, with five quality tiers labeled low, medium, high, xhigh, and max, plus 1k, 2k, and 4k output. That is the same routing discipline I wrote about in the effort-dial piece: the default is rarely the right setting, and max quietly multiplies your bill. Second, reviewers still flag a slightly pasted-in look on foreground subjects, and the default ChatGPT output lands around 1672×941, under 1080p, so production work needs an upscale step, per MindStudio’s tested review. It is a thinking model that follows literal instructions well, not an aesthetics engine.
The four tiers, and the job each one owns
The high-volume tier: Nano Banana 2 and Flare. Google’s Nano Banana 2 is the gemini-3.1-flash-image model, priced around $0.045 at 1K, $0.067 at 2K, and $0.101 at 4K per image, and free inside the Gemini app. It is built for throughput, conversational editing, and anything that sits inside the Google stack. Flare competes directly here on latency. If you are generating hundreds or thousands of images programmatically, this is your tier, and the decision between them comes down to which cloud you already run in. I covered Nano Banana 2 in depth when it launched; a year on, it is still the value floor for high volume work.
The precision tier: Sunburst and GPT Image for text and diagrams. When the image has to contain correct text, a legible chart, a product spec, or a UI mockup where the words matter, the thinking models pull ahead. GPT Image 2.5 beat both Nano Banana 2 and Nano Banana Pro on a four-angle character-consistency test, and its literal instruction-following is the reason. This is the tier for technical, typographic, and product-accurate work where a pretty image that says the wrong thing is a failure.
The creative-direction tier: Midjourney. Midjourney moved its default to V8.2 on July 24, 2026, though many users still run V7. It has no traditional public API and sells on subscription, roughly $10 to $120 a month, billed as GPU time rather than per image. That is a deliberate boundary. Midjourney is a human-led creative tool for aesthetic exploration and brand hero art, not a pipeline component. If your job is one striking image a designer will iterate on, it wins. If your job is ten thousand images a server generates overnight, it is the wrong tool at any price.
The sovereign tier: FLUX.2, run on your own hardware. Black Forest Labs’ FLUX.2 is the open-weight quality leader. The klein variant is a 4-billion-parameter model under an Apache 2.0 license that fits in roughly 8GB of VRAM and costs about $0.014 per image hosted, and the dev checkpoint is self-hostable with a commercial license. It supports native 4-megapixel output and multi-reference composition without fine-tuning, and it runs on consumer NVIDIA RTX cards with FP8 or GGUF quantization, as the local-generation guides document. This tier exists for one reason the other three cannot serve, and I will come back to it.
| Tier | Lead models | Rough cost | Access | Best job |
|---|---|---|---|---|
| High volume | Nano Banana 2, GPT Image 2.5 Flare | ~$0.03 to $0.10 / image | Cloud API | Programmatic, high-throughput generation |
| Precision | GPT Image 2.5 Sunburst / GPT Image | Token-based, ~same as Flare | Cloud API | Text, diagrams, product accuracy |
| Creative direction | Midjourney V8.2 | $10 to $120 / month | Subscription, no API | Hero art, human-led exploration |
| Sovereign / open | FLUX.2 (klein, dev) | $0.014 / image or your GPU | Self-host or API | Private data, unlimited volume, control |
The metric that actually decides it
Sticker price per call is the number everyone compares and the wrong one to optimize. The number that hits your budget is cost per usable image: the per-call price divided by your acceptance rate, multiplied by your retries. A model at $0.03 a call that you accept one in four times costs you $0.12 a usable image before anyone has looked at it. A model at $0.05 that lands on the first try costs less.
This is the same trap I described in the cost-map piece for text, and the same one behind the end of tokenmaxxing: the cheapest token is not the cheapest outcome. The only way to know your real cost is to take 12 to 20 prompts from actual upcoming work, run them through two or three candidates, and track acceptance rate, retries, and the upscale step. Leaderboards rank generic prompts. Your pipeline runs yours.
In two decades running IT operations, the pattern holds everywhere: the line item that kills a budget is never the unit price on the quote, it is the rework nobody modeled. Image generation is no different. Measure the retries.
When self-hosting is not optional
The sovereign tier is the one most buying guides treat as a hobbyist footnote, and for a large class of real work it is the only correct answer. If the images involve customer data, unreleased product photography, medical or legal material, internal documents, or anything you are contractually or legally barred from sending to a third-party endpoint, then quality rankings are irrelevant. The decision was made for you the moment the data became sensitive. You run FLUX.2 on your own hardware, or you do not do the job.
That is a data-residency decision, not an aesthetics one, and it is the reason the open-weight tier matters more than its benchmark scores suggest. It is the same argument for running your own language model that I made in the self-hosting guide: control and residency are features you cannot buy back from an API. For regulated buyers, FLUX.2 [dev] on an RTX box is not the budget option. It is the compliant one.
How to route, in one pass
Start every image task with one question: can this leave your building? If no, you are on FLUX.2, self-hosted, and the rest of this does not apply. If yes, ask whether a human is directing one hero image (Midjourney) or a machine is generating many (Nano Banana 2 or Flare). Then ask whether the image has to carry correct text or precise detail (GPT Image 2.5 Sunburst) or just has to look right (the fast tier). Four questions, four answers.
The field stopped being a horse race this year. GPT Image 2.5 pricing its fast and slow models identically is the clearest signal yet: you are not choosing a winner, you are routing a workload. The buyers who win in 2026 are the ones who already know their acceptance rate per tier, the same way the video-generation field sorted out earlier into tools for different jobs rather than one tool to rule them. Run the test on your own prompts, write down the numbers, and the routing picks itself.
Sources: OpenAI GPT-Image-2.5 API docs, WaveSpeed GPT Image 2.5 breakdown, MindStudio GPT Image 2.5 review, AI Free API model comparison, Digital Applied local image generation 2026.