GPT Image 2.5 Pricing: Cost per Image, Flare vs Sunburst

GPT Image 2.5 is billed per token, which is why "how much does an image cost" is hard to answer from the price list alone: you need to know how many output tokens an image is. We do, for one size at every tier, because the API reports them. This page turns the token counts from our benchmark into per-image prices at OpenAI's list rate, and settles the question people most often get wrong. No, Sunburst does not cost more than Flare.

The rates

Both models are listed at the same per-token prices:

For a text-to-image request the image-output line is the whole bill in practice. In our runs the prompt itself was 39–104 text-input tokens: a few hundredths of a cent at $5 per million.

Cost per image, 1024×1024

Output tokens from our benchmark — the same product-shot prompt at every tier, both models — multiplied by $30 per million:

Read the column top to bottom: max is 36× the tokens of low for the same prompt at the same size, and high is 9×. The tier moves your bill by more than an order of magnitude; the model does not move it at all.

Flare and Sunburst cost the same, to the token

"Same price list" would be a weak claim on its own; a model could still consume more tokens per image. It does not. In all five tiers, gpt-image-2.5-flare and gpt-image-2.5-sunburst returned identical output-token counts for the same request: 196, 439, 1,756, 3,122 and 7,024. Our four other prompt pairs at high — a portrait, a sprite sheet, and a bookshop with two different exclusion lists — came back at 1,756 tokens on both models every time.

Flare10 s

gpt-image-2.5-flare

Sunburst16 s

gpt-image-2.5-sunburst

So the choice between the two models is never a budget decision. It is a time decision (see the speed page) and a quality decision; the single-model pages for Flare and Sunburst go through what each one won and lost.

What changes the price

• Quality tier, above all. Five explicit tiers, a 36× spread.
• Size. Larger outputs are more output tokens. We have only measured 1024×1024 so far, so we do not publish per-size prices; when we do, expect both models to move together, as they did across tiers.
• Reference images, if you send them — the character sheet in the consistent-characters workflow, say: image-input tokens at $8 per million, with cached input cheaper.
• auto quality. Under auto each model chooses its own effort, so the token count — and the bill — is no longer fixed by you. Set the tier explicitly if you want a predictable price.

Not the model.

List price versus what a gateway charges

Every price on this site is OpenAI's list price: output tokens × $30 per million. It is not what we paid. Round one ran through a third-party API gateway whose per-token rate is its own, not OpenAI's; its invoice for all 18 images was $0.9432, while the same output tokens at list price come to about $1.17. If you call OpenAI directly, the table above is your price. If you buy through a reseller, use the reseller's rate. The token counts still apply either way, because the model reports them the same way regardless of who bills you.

The same distinction applies here: the studio on this site charges credits, priced on the pricing page, not the API rate. It shows the list price next to each image so you can see what the same request would cost direct.

Work out your own price

1. Send one request at the tier and size you intend to ship.
2. Read output_tokens from the response.
3. Multiply by $0.00003, thirty dollars per million.
4. Add the text-input tokens at $5 per million if you want to be exact; for a prompt under a few hundred tokens the addition is well under a cent.

The same output-token count will apply to both models, so one request tells you the price on either. The studio shows output tokens and list price under every image it generates, on either model, and charges the same credits for a given tier whichever model you pick.

FAQ

Is gpt-image-2.5-sunburst more expensive than gpt-image-2.5-flare? No. Same per-token rates, and in our measurements the same token counts per image at every tier.

How much does one GPT Image 2.5 image cost? At 1024×1024 and list price: about $0.006 at low, $0.013 at medium, $0.053 at high, $0.094 at xhigh and $0.21 at max. Other sizes cost more or less in proportion to their output tokens.

Which tier is the best value? The lowest one that passes your own quality bar. Every step up multiplies the price: about 2.2× from low to medium, 4× from medium to high, then 1.8× and 2.3×. All of our playbook tests — a photorealistic portrait, a 16-frame sprite sheet, a detailed interior — were run at high.

Does a longer prompt cost more? Fractionally. Text input is billed at $5 per million tokens, so ten extra words is a rounding error next to a single output image.

I am on gpt-image-1. What will 2.5 cost me? Less per token: OpenAI lists gpt-image-1 at $40 per million image-output tokens ($10 per million image input) and gpt-image-1.5 at $32, against $30 on either 2.5 model. Per image depends on the output tokens each model returns for the same request, so compare output_tokens from both at the tier you ship. The migration guide lists what else changes before the 2026-10-23 shutdown.

Why do some sites say Sunburst is the more expensive model? Because it is the quality-optimised one, and quality usually costs more. Here it does not: the price list is identical and the token counts were identical in every one of our runs.

Sources

• OpenAI API pricing — per-token rates for both 2.5 models and the older generations, checked 2026-09-20
• Our runs: 2026-09-12, 1024×1024, one run per cell; output tokens as reported by the API for each job