Skip to main content
Generate images from a text prompt with OpenAI GPT Image or Google Gemini. POST /v1/images/generations accepts OpenAI’s Images API shape and returns base64-encoded image bytes — the model never returns URLs. Use GET /v1/models for current availability and rates. Image generators have output_modality: "image" and output_modalities: ["image"].

​
Generate an image

Response (abridged)
Decode data[0].b64_json to get the image file. A revised_prompt field accompanies an image when the provider rewrote your prompt before generating. For Gemini, replace model with a Google model above. The response includes data[0].mime_type, either image/png or image/jpeg, identifying the actual encoded bytes. Use that MIME type when displaying or saving the image.

​
OpenAI options

​
Gemini options

Gemini supports one image per request (n: 1). Omit size or set it to auto for a square 1K image; supported explicit square sizes are listed above. quality and background can be omitted or set to auto. Omit output_format: Gemini selects PNG or JPEG natively, and returns the format in data[].mime_type. Explicit output_format, output_compression, and user values are unsupported. Unsupported settings return 400 before a provider request is sent. This endpoint currently accepts text prompts; reference-image editing is not exposed.

​
Billing

Image output is billed per token like any other model — text input tokens plus image output tokens, at the rates in Models & pricing. Gemini additionally reports text and thinking output tokens, which use its separate text_output_per_mtok_usd rate instead of the image output rate. usage.output_tokens_details separates image_tokens and text_tokens (including thinking); their sum is usage.output_tokens. Internal thinking does not make an image generator a text-output model. The cost is settled when generation finishes and appears in GET /v1/usage and the request log. Two guardrails to know:
  • Model and endpoint must match: sending an image model to a text endpoint (/v1/chat/completions, etc.) fails with 400 invalid_request_error, as does sending a text model to the images endpoint.
  • Image generation always draws on the prepaid balance — there is no BYOK path for it, so an empty balance returns 402 even with a provider credential configured.