What Is GPT Image 2.5? Sunburst and Flare

Your feed filled up overnight. "The new king of AI images." "OpenAI just fixed image editing." Somewhere in that noise are two model names you've never seen before — Sunburst and Flare — a claim about 50% lower latency, and a pricing page that looks identical for both models even though one is explicitly slower. If you're trying to work out whether this actually changes anything for you, the hype is useless.

Here's what GPT Image 2.5 is, stated plainly and sourced entirely to OpenAI's own announcement, model pages, and pricing tables rather than to launch-day commentary. Where OpenAI has published a number, this article quotes it. Where OpenAI has not, this article says so instead of filling the gap with an estimate.

The short answer

GPT Image 2.5 is OpenAI's image generation and editing model family, released on September 8, 2026. It ships in two places at once: as the image model inside ChatGPT (branded "ChatGPT Images 2.5"), and as two separately named models in the API — GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.

The headline improvements OpenAI claims over the previous generation:

• Up to 50% lower generation latency compared with Images 2.0
• Better preservation of subjects from your reference photos
• Edits that change only what you asked for, leaving the rest intact
• Multi-turn editing that stays consistent instead of degrading as the conversation goes on
• Better handling of complex layouts, including transparent backgrounds

For scale, OpenAI states that people create more than 3 billion images per week across ChatGPT Images and the GPT-Image API models. This is not a niche release.

The pain point this article solves

Launch coverage tells you a model is "better." It almost never tells you the thing you actually have to decide: which of the two models to call, whether your existing cost estimates still hold, and whether the free ChatGPT tier gets it. Those three answers are buried in three different OpenAI pages. They're collected here.

Sunburst vs Flare: what the two names mean

This is the single most confusing part of the release, because the naming gives you no hint about the tradeoff.

GPT-Image-2.5 Flare is marked Default on OpenAI's model page. OpenAI describes it as "the default choice for most applications, delivering higher-quality images than GPT-Image-2 at 50% lower latency," and recommends it for creator and social content, product experiences, visual search, rapid prototyping, and high-volume generation.

GPT-Image-2.5 Sunburst is the precision tier. OpenAI says it "offers an extra level of precision for detailed creative work with longer generation times" and is "built for premium visual workflows that benefit from tighter control across edits" — production-ready campaign creative, polished product imagery.

So the tradeoff is speed versus editing precision, and Flare is the one OpenAI itself points most developers toward. We break the choice down case by case in Sunburst vs Flare: which model to actually call.

The part almost every launch article gets wrong: Sunburst is not the "expensive" tier. On OpenAI's pricing page the two models carry identical token rates — $8 per million image input tokens, $2 per million cached image input tokens, $30 per million image output tokens, $5 per million text input tokens, $1.25 per million cached text input tokens. Same numbers, both models, line for line.

That does not mean the same cost per image. OpenAI's own guide is explicit about the distinction: "Equal token rates don't mean equal cost per image: token consumption can differ by model and quality setting." Your bill moves with how many tokens a given model and quality setting actually consumes, not with a per-model markup — see what a GPT Image 2.5 image actually costs for the full arithmetic.

What's new inside ChatGPT

If you use ChatGPT rather than the API, the model change came with four product features:

• Sketch — draw directly in ChatGPT and use the drawing as a visual guide for the final image. Type @Sketch to start one.
• Templates — starting points for common formats like posters and merch, so you're not staring at an empty prompt box.
• Comments on images — place a comment directly on a spot in the image to drive a focused edit.
• Shareable prompts — when you share an image, you can include the prompt that produced it so someone else can run the same idea on their own photos.

Availability: OpenAI states that "Images 2.5 is rolling out today to ChatGPT, ChatGPT Work, and Codex users across all tiers on desktop, mobile, and web." All tiers includes the free tier. What OpenAI has not published is a specific number of free generations per day — so treat any hard figure you see quoted online as unofficial. More on that in is GPT Image 2.5 free, and what you really get.

The quality settings changed, and old cost math breaks

This is the detail most likely to bite an existing integration.

Both 2.5 models support six quality settings: low, medium, high, xhigh, max, and auto. Earlier GPT Image models supported quality settings only up to high. So xhigh and max are genuinely new rungs on the ladder, and auto is the default.

The trap sits right next to it. OpenAI's Flare model page carries this warning: "Token rates match GPT Image 2. The GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption."

Read that twice if you're carrying a spreadsheet over from GPT Image 2. The prices are unchanged. The token counts are not, and OpenAI is telling you the old estimator no longer applies. The correct move is to read the usage field on your own API responses at the sizes and quality settings you actually ship, and rebuild the estimate from real numbers.

Sizes, formats, and the limits worth knowing

From OpenAI's image generation guide:

• Recommended sizes: 1024x1024 (square), 1536x1024 (landscape), 1024x1536 (portrait) — the full rule set is in every supported GPT Image 2.5 image size
• Custom dimensions are supported as WIDTHxHEIGHT strings, with real constraints: width and height must be multiples of 16, the aspect ratio must fall between 1:3 and 3:1, neither edge may exceed 3840 pixels, and total pixel count must land between 655,360 and 8,294,400 (4K). Resolutions above 2560x1440 are flagged experimental.
• Transparent backgrounds: set background: "transparent" with output_format of png or webp.
• Formats: png by default, plus jpeg and webp with an output_compression value from 0-100. OpenAI notes jpeg is faster than png, so reach for it when latency matters.

And the limitations OpenAI publishes rather than hides: complex prompts may take up to 2 minutes; text placement and clarity can still miss; recurring characters or brand elements may drift across generations; and precise placement in layout-sensitive compositions remains hard. Those are the same four weak spots that have dogged every generation of this model family, and 2.5 narrows them rather than closing them.

Try the workflow before you rewire anything

If your goal is to see whether a newer image model actually improves your output — not a demo reel's — the fastest test is to run the same prompt and the same reference photo through more than one model and compare.

That's what Vogoo's GPT Image 2.5 studio is for. It runs both Flare and Sunburst in the browser — you pick the tier per render, up to 4K, with the credit cost shown before you commit — and there are no API keys, no organization verification, and no code. If you'd rather line it up against other models first, start on text to image for generation, or image to image when you want to edit a photo you already have.

Where GPT Image 2.5 sits against the rest of the field

Honest positioning matters more than a victory lap, so here is the shape of the market as of this writing:

• Against its predecessor: faster, more precise on targeted edits, more stable across multi-turn editing, with two new quality rungs. Same token rates, different token consumption — the migration hazards are in GPT Image 2.5 vs 2.
• Against Google's Nano Banana line: Google prices images at a fixed token count per resolution — a 1024x1024 image from Nano Banana 2 is 1120 tokens, or $0.067. OpenAI does not publish an equivalent fixed per-image figure, because consumption varies with model and quality. If you need a predictable unit cost before you ship, that difference matters more than any benchmark chart. Full breakdown: GPT Image 2.5 vs Nano Banana.
• Against Microsoft's MAI line: Microsoft has already moved past MAI-Image-2.5 to MAI-Image-2.6, so any comparison written against 2.5 is describing a superseded model.

None of that makes one model correct for everyone. It makes the choice depend on whether you're optimizing for editing precision, unit-cost predictability, or raw speed.

FAQ

When was GPT Image 2.5 released? September 8, 2026, in both ChatGPT and the API.

Is GPT Image 2.5 free? Inside ChatGPT, yes — OpenAI says it rolled out to ChatGPT, ChatGPT Work, and Codex users across all tiers. OpenAI has not published a per-day free generation count. The API is paid, billed by tokens.

What are the model IDs? gpt-image-2.5-sunburst and gpt-image-2.5-flare. Developers working inside Codex get the model there too — see GPT Image 2.5 in Codex.

Which one should I use? Flare unless you have a specific reason not to — OpenAI marks it Default and recommends it for most applications. Move to Sunburst when editing precision on premium creative work matters more than latency.

Does GPT Image 2.5 watermark images? OpenAI states it uses C2PA metadata and invisible watermarking to help identify images made with its tools.

Can I still use my GPT Image 2 cost spreadsheet? No. The rates match, but OpenAI explicitly says the GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption.

The honest bottom line

GPT Image 2.5 is a real step forward on the two things that actually break creative workflows — edits that wreck the parts you wanted kept, and quality that degrades as a conversation goes on. It is not a step forward on text rendering or precise layout control, and OpenAI says as much in its own limitations list.

If you build on the API, the action item is narrow and concrete: default to Flare, and re-measure your token consumption instead of trusting last month's spreadsheet. Our GPT Image 2.5 API setup guide covers the verification gate and every output parameter, and running your own benchmark covers how to measure it. If you create in a browser, the action item is even simpler — run your real prompt through it and compare the output to whatever you're using now. Run your brief through GPT Image 2.5 in Vogoo and see what Flare and Sunburst each do with it before you commit to either.

Written by the Vogoo team, September 10, 2026. Every figure above is quoted from OpenAI's launch announcement, model pages, and public pricing tables as published on that date; model pricing and availability change frequently, so verify against the linked sources before making a budgeting decision.

Sources

• Introducing ChatGPT Images 2.5 — OpenAI — launch date, latency claim, ChatGPT feature set, tier availability, API model positioning
• Image generation guide — OpenAI API docs — sizes, quality settings, custom dimension constraints, limitations, cost mechanics
• GPT-Image-2.5 Flare model page — OpenAI API docs — Default designation, quality options, calculator warning
• OpenAI API pricing — token rates for both 2.5 models
• Gemini API pricing — Google — Nano Banana per-image figures used in the comparison