GPT Image 2.5 API Pricing: Cost per Image and Budget Examples

GPT Image 2.5 has two official API models, Sunburst and Flare, with the same Standard token rates. For a 1024×1024 image, OpenAI's calculator estimates $0.00588 at low quality, $0.01317 at medium, and $0.05268 at high. These are image output estimates; text prompts, reference images, and other applicable charges increase the request total. The figures below were checked on September 9, 2026, using the calculator's explicit GPT Image 2.5 (Sunburst and Flare) setting. Official calculator and cost guide

For a quick planning number, 1,000 square images at high quality have an estimated output cost of $52.68. If each request also consumes 1,000 uncached text input tokens, the estimate becomes $57.68, before reference images, streamed previews, or additional attempts. That input count is an assumption for the calculation, not a measurement of a typical prompt.

Estimated cost per image at each quality setting

The following values come from operating OpenAI's official calculator with GPT Image 2.5 selected and switching through all five explicit quality settings. They apply to the stated dimensions and represent image output only, not a fixed fee for a completed request. OpenAI image generation calculator

The landscape values correspond to 158, 343, 1,372, 2,459, and 5,488 output tokens, respectively. Multiplying any output token count by $30 per million produces its dollar estimate. For example, 1,756 × $30 ÷ 1,000,000 = $0.05268.

The landscape estimates are lower than the square estimates in these calculator results, despite the landscape image having more pixels. Treat each size and quality combination as its own estimate; scaling the square price by pixel area would give the wrong answer here. This observation does not establish that every wider image or custom size is cheaper.

The calculator groups Sunburst and Flare together, but OpenAI cautions that actual consumption can vary by model, prompt, and quality. We did not measure API billing for these examples. For a repeatable budget, specify size and quality instead of leaving them at auto, then compare the returned usage with your estimate.

Official token rates and model selection

Both gpt-image-2.5-sunburst and gpt-image-2.5-flare have the following Standard rates in USD per one million tokens, checked September 9, 2026. OpenAI image generation pricing

These models return images; there is no text output tariff in their image model pricing rows. The lower cached rates apply to tokens reported as cached, not every repeated prompt or reference image.

OpenAI positions Sunburst for precise editing and Flare for fast everyday image generation. Their names do not designate different price tiers. Choose based on the work you need to complete, then measure how much usage each consumes and how often its results are usable. Equal rates per token do not guarantee equal cost per accepted image.

Use a documented model ID in your implementation. For direct Image API requests, that means gpt-image-2.5-sunburst or gpt-image-2.5-flare; do not substitute an assumed bare gpt-image-2.5 alias. In the Responses API, the top-level model is a supported mainline model, while the image generation tool's model selects the image model. The mainline model's tokens add a separate cost. Image generation API guide

The current Batch image pricing table we checked does not list these two GPT Image 2.5 IDs. Do not build a budget around an assumed 50% Batch discount. Recheck the applicable pricing table when choosing a processing mode. These figures are API usage prices, not ChatGPT subscription prices.

Add the costs that a per-image table leaves out

For the image model itself, calculate each request using disjoint token categories:

text
Image model cost in USD = (
    5.00 × uncached text input tokens
  + 1.25 × cached text input tokens
  + 8.00 × uncached image input tokens
  + 2.00 × cached image input tokens
  + 30.00 × image output tokens
) / 1,000,000

Do not add a total input token count to its text and image breakdown. Likewise, if an input total includes cached tokens, subtract those before calculating the uncached portion. The formula applies the official rates; actual quantities should come from the API's reported usage.

Consider a hypothetical square, high-quality generation with 1,000 uncached text input tokens and the calculator's 1,756 output tokens. Text costs $0.005 and output costs $0.05268, giving $0.05768 per request.

Now suppose an editing request uses the same text and output counts plus 2,000 uncached image input tokens. The reference image input adds $0.016, bringing the estimate to $0.07368. If those same 2,000 image input tokens are instead reported as cached, their charge is $0.004 and the total becomes $0.06168. These assumed token counts illustrate accounting; they do not predict the tokenization of a specific uploaded image.

Streaming adds another consideration. OpenAI states that each partial image adds 100 image output tokens. At the 2.5 output rate, that is $0.003 per partial image, or $0.009 for three. Include that increment when forecasting from an output-only calculator estimate. When reconciling actual usage, do not add it again if those tokens are already included in the reported total. Streaming and image generation costs

Budget for images you can actually use

A production budget should distinguish generated images from accepted images. A result can be a successful API response and still need another attempt because the composition, editing precision, or other requirements were missed.

Using the $0.05768 request estimate above, suppose a job needs 1,000 accepted images and you assume that 80% of completed generations will be accepted:

text
Expected generations = 1,000 / 0.80 = 1,250
Estimated total = 1,250 × $0.05768 = $72.10
Cost per accepted image = $72.10 / 1,000 = $0.07210

The 80% acceptance rate is a planning assumption, not a measured property of either model. This example holds the request's text input and output consumption constant and excludes reference image inputs, streaming partials, and Responses mainline usage. It also does not assume how a provider charges for failed requests.

Before committing to a monthly volume, run a representative set of your own generation or editing tasks. Record the model ID, dimensions, quality, reported token usage, total expense, and number of results you accepted. Compare:

text
Observed cost per accepted image = total expense / accepted images

A lower quality setting reduces the calculator estimate, but may require more attempts for your particular task. Conversely, paying for max does not automatically make the overall workflow cheaper. The useful comparison is the expense required to deliver an acceptable result under consistent acceptance criteria.

For example, 10,000 square outputs at medium have a calculator output estimate of $131.70. The same output count at high is $526.80. Those numbers help set an initial range; measured inputs and acceptance rates determine which setting offers the better production budget.

Avoid mixing GPT Image 2 and 2.5 prices

Some familiar square-image figures—roughly $0.006 at low, $0.053 at medium, and $0.211 at high—belong to GPT Image 2, in the guide's earlier models section. They should not be relabeled as GPT Image 2.5 prices.

The difference is especially easy to miss at medium: the explicit GPT Image 2.5 calculator selection gives $0.01317 for a 1024×1024 output. Similar-looking prices at different quality labels do not prove that those settings have equivalent visual quality. A claim of a particular quality-adjusted saving would require an actual comparison.

There is also a documentation detail to watch: the 2.5 model pages warn against using the GPT Image 2 calculator for Sunburst or Flare. On September 9, the guide's calculator did offer a separate GPT Image 2.5 (Sunburst and Flare) selection, which is what this article used. Check the selected model before copying a quote.

If you are maintaining an older integration, our GPT Image 2 pricing guide covers that model's separate pricing context, and the GPT Image 2 API guide covers its implementation. For a new 2.5 budget, start with the explicit 2.5 calculator setting, add the applicable input and workflow costs, and replace your assumptions with observed usage as soon as you have representative requests.