Skip to main content

​
Short Answer

This is normal. A 4K image at high quality is inherently expensive in GPT Image, and even a single output image can consume a large number of output_tokens. Image output tokens are not calculated simply from “one output file” or as a fixed linear ratio of total pixels. They are mainly affected by:
  1. quality: low, medium, high, or auto
  2. Output dimensions and aspect ratio
  3. Generation count n
  4. The model’s internal canvas partitioning and image complexity
usage.output_tokens represents the image tokens used to generate the output image, not reference-image input. Reference images are reported separately under usage.input_tokens_details.image_tokens.

​
Why Can One Image Use So Many Tokens?

“One image” describes the number of returned results, not the amount of generation work. The model must generate the entire canvas in its internal image-representation space. Higher quality and larger resolution generally require more image computation and output tokens. For example, both requests return one image:
The second request still returns only one image, but it uses a 4K landscape canvas and the high-quality tier. A much larger output-token count is expected.

​
Four Main Cost Factors

​
1. The quality Parameter

quality is usually the most visible cost variable:
With quality: "auto" or no explicit quality, the model may select a different tier based on the prompt. Calls with identical dimensions and reference images can therefore differ by several times in output_tokens. For predictable budgeting, explicitly set low, medium, or high.
The project documentation includes a real example: three requests each used 1061 input tokens, while their output counts were 1286, 5146, and 1287 tokens. The middle call automatically selected a higher quality level and cost about 3.5 times as much as the other two.

​
2. Output Dimensions

Higher resolution generally means a larger internal canvas and more output tokens. A 4K high request can cost far more than a 1K low request. However, the following formula cannot predict the exact result:
Pixel count is useful only for rough budgeting. The model determines the actual output_tokens during generation, and the response’s usage.output_tokens is the source of truth.

​
3. Aspect Ratio and Internal Canvas Partitioning

Output tokens also depend on how the internal canvas is tiled, scaled, and covered. Token usage is therefore not always strictly monotonic with the final pixel count. At the same quality tier, a larger non-square image can sometimes consume fewer output tokens than a smaller or more square image. This is not contradictory: the model uses discrete canvas or tiling rules rather than billing each final pixel independently.
When comparing dimensions, consider both quality and aspect ratio. Do not compare only the “4K” or “2K” label or total pixel count.

​
4. Generation Count n

As a general rule, generating more images increases total output tokens. N generated images carry roughly N sets of output cost. However, the current gpt-image-2 endpoint supports only n=1. To generate multiple images, send multiple independent requests; each request is billed separately for input and output tokens. Whether another image model supports n>1 depends on that model’s documentation.

​
Which Tokens Are Affected by Reference Images?

Reference-image count primarily affects input image tokens, not output image tokens: gpt-image-2 processes reference images at high fidelity. More reference images increase input image tokens approximately linearly, while the final output’s output_tokens are still mainly determined by output quality, dimensions, aspect ratio, and the model’s internal generation process. If the billing record explicitly attributes the large count to “image output,” it should not be blamed on the number of references. Reference-image cost should appear separately in the input-image token field.

​
How to Calculate the Real Cost

Using the current gpt-image-2 billing structure:
Example response:
This means:
  • 16 tokens came from prompt text
  • 1024 tokens came from the reference image
  • 5146 tokens came from the generated image
  • Only one image was returned, but its quality and canvas required 5146 output image tokens
There is no universal fixed token count per 2K or 4K image across all dimensions and content. Budget tables are estimates only. Final billing should use the actual usage values returned by the API or shown in the console logs.

​
How to Reduce Token Usage

1

Set an explicit quality tier

Avoid auto. Pass low, medium, or high explicitly so the model does not unexpectedly select a more expensive tier.
2

Avoid unnecessary resolution

Use 1K or 2K for previews and internal review, then generate 4K high only for final delivery.
3

Choose the required aspect ratio

Use the canvas your workload actually needs instead of increasing dimensions only to appear more detailed.
4

Control output count

Multiple candidates increase total output cost approximately linearly. Validate prompts with a small batch before scaling up.
5

Record usage fields

Store each request’s size, quality, output_tokens, and cost to build a cost baseline from real production data.