Source: OpenAI Image Generation Guide

OpenAI API guide to GPT Image models: Image API vs Responses image_generation tool, edits, masks, markup, sizes, quality, streaming partials, pricing and limitations.

Contents
  • What: official API documentation (Image generation guide)
  • Author/publisher: OpenAI
  • Date: living document; read 2026-11-14
  • Original: Image generation

Summary

Describes OpenAI’s GPT Image family and the two ways to call it: the Image API (Generations and Edits endpoints) and the Responses API’s built-in image_generation tool, which adds multi-turn editing inside a conversation.

Key claims

  • Models named (as of 2026-11): gpt-image-2.5-sunburst (editing precision), gpt-image-2.5-flare (fast everyday generation), GPT-Image-2.6 variants, and gpt-image-2. Organization verification may be required.
  • Responses API: a mainline model is chosen at top level and the image model in the tool; action is auto, generate or edit; the mainline model may rewrite the prompt (revised_prompt). Multi-turn editing uses previous_response_id.
  • Edits accept multiple reference images. Masks are prompt-guided, so the model may not follow the mask exactly. “Markup” edits send a transparent annotation layer.
  • Sizes: 1024x1024, 1536x1024, 1024x1536 or custom (multiples of 16, aspect 1:3 to 3:1, edge at most 3840 px); above 2560x1440 is experimental.
  • quality: low/medium/high/auto on gpt-image-2; GPT-Image-2.6 variants add xhigh and max and require a total pixel count of 655,360-8,294,400; separate complexity setting (low/medium/high).
  • Output formats png (default), jpeg, webp, plus svg and animated gif on newer models; transparent background requires png or webp; JPEG is faster than PNG.
  • Streaming: partial_images 0-3; each partial adds 100 output tokens.
  • Pricing is per token, e.g. image output $30 per million tokens for GPT-Image-2.5/2.6; gpt-image-2 at 1024x1024 is about $0.006 / $0.053 / $0.211 for low / medium / high.
  • Complex prompts can take up to 2 minutes.
  • Stated limitations: latency, text rendering precision, consistency of recurring characters or brand elements, precise layout placement.
  • moderation is auto (default) or low; blocked requests return moderation_blocked.

Feeds

Multimodal models

Suggest changes →