GPT Image 2.6 Pricing: Sunburst vs Flare, Which to Use?

How much does GPT Image 2.6 cost, and which version should you use?

GPT Image 2.6 is OpenAI's newest image model. It comes in two versions: Sunburst, built for higher quality, and Flare, built for speed. The one price reported so far is $8 per million input tokens, and OpenAI lists separate input and output pricing. The reports do not say which version that rate applies to, or what the output rate is. Use Sunburst for final, customer-facing images. Use Flare for drafts, volume work and quick iterations.

What is GPT Image 2.6?

GPT Image 2.6 is OpenAI's latest image generation model. You describe a picture in plain language, and it produces one. The 2.6 release adds two things people have asked for: faster generation and animated GIFs.

The speed gain is reported at about 50%. That is a rough figure from launch coverage, not a benchmark you can rely on for planning. Your real speed will depend on image size, quality settings and how busy the service is.

One note on context. The model was covered in an AI news roundup in which the presenter described OpenAI as possibly having the strongest image model among the big labs. Other image models, including ByteDance's Seedream 5.5, were also praised in the same roundup. Treat "best" claims as opinions and test the model on your own images.

What is the difference between Sunburst and Flare?

OpenAI split the model into two variants that trade quality against speed.

The "likely costlier" and "may look less refined" entries are inferences from how the two are described, not published figures. Fast and cheap variants of models usually give up some detail, and that is the sensible assumption here until you test it.

The split is useful because most image work happens in two phases. First you explore: ten versions of a concept, different layouts, different moods. Then you finish: one image that needs to look right. Flare fits the first phase and Sunburst fits the second.

How does GPT Image 2.6 pricing work?

OpenAI prices the model with separate input and output rates, and the one rate reported is $8 per million input tokens.

If "tokens" sounds odd for pictures, the confusion is common. A token is a unit of data the model reads or writes. Text models count chunks of words. OpenAI's earlier GPT image models in the API were also billed by tokens: your text prompt counts as input, any reference image you upload counts as input, and the image the model draws counts as output. So token billing for images is an established pattern, not a new quirk.

Here is what that means in practice:

1. Your prompt costs a little. A short description is a small number of tokens.
2. Reference images add input cost. If you upload a product photo and ask for edits, that image is counted.
3. The generated image is the output. Output is usually the larger part of the bill, and larger or higher-quality images use more tokens.
4. The rate is per million tokens. To get a per-image cost, multiply the tokens used by the rate and divide by a million.

As arithmetic only: 1,000 input tokens at $8 per million is $0.008. That is an illustration of the formula, not a measurement of what an image costs. The output rate, which probably drives most of the cost, is the number to look up.

How to find your real cost per image

Do not guess. Run a small test:

1. Generate 20 images with Flare and 20 with Sunburst, using prompts typical for your work.
2. Check the usage numbers in the API dashboard (or your admin view) after the batch.
3. Divide total spend by 20 for each variant.
4. Compare the per-image cost against how many images you usually discard before you accept one.

That last step matters. A cheaper model that needs five tries can cost more than a pricier one that gets it right on the second.

Can GPT Image 2.6 make GIFs?

Yes. GIF generation is one of the new features, and the launch demos showed two example GIFs. You describe a short looping animation, and the model produces an animated file instead of a still image.

Practical uses for people at work:

• A looping animation for a slide deck or internal newsletter
• A simple animated header for an email or help-center article
• A reaction or celebration GIF for a team chat
• A before-and-after or step-by-step visual that benefits from motion

Keep expectations measured. The sources show the feature exists, but they do not detail clip length, frame rate, file size limits or how the GIF is priced. Test a few before you promise a stakeholder a library of animated assets. Also check file size, because large GIFs load slowly in email and web pages.

Where can you use GPT Image 2.6?

The model is available in three places:

• ChatGPT. The easiest route. You type a request and get an image. No setup, no billing math beyond your plan.
• Codex. OpenAI's coding tool, useful if you are building something and want images generated as part of the work.
• The API. For teams that want to generate images inside their own software, on demand, billed by usage. This is where the token pricing applies directly.

If you are a regular ChatGPT user, the pricing question mostly doesn't apply to you. You pay your subscription and use the model within its limits. The per-token rates matter if you are building a workflow, generating at volume, or running a product on top of the API.

Which version should you choose?

Use this as a starting rule and adjust after your own test.

A good habit is to draft on Flare, pick the direction, then regenerate the chosen concept on Sunburst. You spend the higher rate only on images that matter.

What this means for you

If you are a marketer or communicator: the speed gain makes iteration cheaper in time. Ask for many variations, choose one, and finish it on Sunburst. Add GIFs to your toolkit for decks and emails, but check file sizes.

If you are a manager or team lead: set a simple rule for which variant your team uses, so costs don't drift. Run the 20-image test above and share the per-image numbers.

If you are a developer or analyst building workflows: model your costs from real usage, not from the headline $8 rate. Track input and output tokens separately, because they have separate prices.

If you only use ChatGPT: you don't need to manage tokens. Try both quality levels if they are offered to you and note which one you prefer for which task.

For everyone: confirm current prices and limits on OpenAI's official pricing page. Launch-day numbers and later numbers can differ, and this article relies on early coverage.

Can you build your own image request tool instead of paying per seat?

If your team generates images regularly, a shared tool can beat everyone using personal accounts. With Remy, you describe an internal app in plain English, such as an image request form where teammates submit a brief, the app generates options, saves approved images to a shared library, and logs what each request cost. Remy's model catalog covers nearly 400 models, including image and video models, and AI usage is billed at cost. The catalog changes daily, so check whether a specific model, such as GPT Image 2.6, is listed before you plan around it.