Multi-turn GPT Image 2.5 editing runs only through OpenAI's Responses API

OpenAI's image generation guide says the API lets you generate and edit images from text prompts using gpt-image-2.5-sunburst and gpt-image-2.5-flare, and to choose Sunburst for workflows where editing precision matters most and Flare for fast, high-quality everyday image generation.

OpenAI calls Flare "the default choice for most applications".

OpenAI describes Sunburst as "built for premium visual workflows that benefit from tighter control across edits", and its launch announcement says Sunburst's better precision comes with "longer generation times".

With the Image API you set model to gpt-image-2.5-sunburst or gpt-image-2.5-flare directly; with the Responses API you select a supported mainline model at the top level and specify the image model in the image generation tool's model field.

The docs' Image API example calls client.images.generate(model="gpt-image-2.5-sunburst", prompt=prompt) and reads the image from result.data[0].b64_json before base64-decoding it to bytes.

The docs' Responses API example calls client.responses.create with model="gpt-6-astra" and tools=[{"type": "image_generation", "model": "gpt-image-2.5-sunburst"}], then collects image data by filtering response.output for outputs whose type is image_generation_call.

The New Stack reports that OpenAI launched GPT Images 2.5 this week, promising better results for changing one part of an image without disturbing the rest, including edits to a single element such as a product, background or piece of copy while preserving the surrounding scene.

Both APIs let you customize output by adjusting quality, size, format, and compression.

The New Stack reports that OpenAI lists the same token rates for Flare and Sunburst: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens.

OpenAI states: "Token rates match GPT Image 2. The GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption."

The New Stack reports that although image generation pricing is based on token counts, OpenAI gives no explicit note on how to estimate GPT-Image-2.5 token consumption.

The guide directs readers to a supported models list for the mainline models that can call the image generation tool.

Compared with the Image API, the Responses API adds multi-turn editing (iterative high fidelity edits by prompting) and flexible inputs, accepting image File IDs as input images, not just bytes. It supports image generation as a built-in tool and accepts image inputs and outputs within context.

Because both models post the same per-token rates and OpenAI publishes no way to estimate GPT Image 2.5 token consumption, the per-image cost difference between Flare and Sunburst has to be measured on real requests rather than computed from the published price list.

At $30 per million image output tokens, image output costs 3 cents per thousand output tokens, so a 1,000-token image costs 3 cents and a 4,000-token image costs 12 cents.

With the Responses API, multi-turn image generation works either by providing image generation call outputs within context (the image ID also works) or by using the previous_response_id parameter.

The optional action parameter on the image generation tool controls whether a supported tool model generates a new image or edits one in the conversation: auto lets the model decide, generate always creates a new image, edit forces editing when an image is in context.

The guide states that if you force action "edit" without providing an image in context, the call will return an error.

The New Stack reports that neither OpenAI's launch announcement nor the model pages for Flare or Sunburst make clear the difference in generation time between the two models.

A Responses API image request names two models, the top-level mainline model and the image model inside the tool, where an Image API request names one.