GPT-Image-2 API Guide: GPT Image 2.5, Responses Tool, Pricing

As of October 10, 2026, the GPT-Image-2 API is live and not deprecated. gpt-image-2 (snapshot gpt-image-2-2026-04-21) still runs on /v1/images/generations, /v1/images/edits and /v1/batch, and it can still be the model behind the Responses image_generation tool.

OpenAI's image generation guide now points new integrations elsewhere, though: "For new integrations, use one of the GPT Image 2.5 models." That means gpt-image-2.5-sunburst when editing precision matters most, or gpt-image-2.5-flare for fast, everyday generation. All three models share the same endpoints, token prices and rate limits. Moving from gpt-image-2 to 2.5 is mostly a change of model string, plus two extra quality levels (xhigh and max).

The models that do need migrating are older ones. gpt-image-1 shuts down on October 23, 2026, and gpt-image-1.5, gpt-image-1-mini and chatgpt-image-latest follow on December 1, 2026.

For a full budget by size and quality, see GPT Image 2.5 API pricing in detail. A compact table of 2 and 2.5 prices sits in the pricing section below.

Quick Answer

Official Evidence

Part of the guide carries the note "The details below apply to earlier models, not Sunburst or Flare." Where a 2.5 behavior isn't documented, the tables below say so instead of assuming it matches gpt-image-2.

GPT Image 2.5 vs gpt-image-2: Which API Parameters Change

The request shape is the same for all three models. What differs is the quality range, the status of transparent backgrounds, and input_fidelity.

Pick Flare when you need speed and volume for everyday images, and Sunburst when an edit has to keep details intact. Staying on gpt-image-2 is reasonable if your prompts are already tuned and nothing in this table blocks you. For the trade-offs with numbers, see GPT Image 2.5 Sunburst vs Flare: Which to Use, by the Numbers and what changed from GPT Image 2 to 2.5.

Choose Image API or Responses First

Minimum Viable Calls

The smallest useful call goes to the Image API and decodes the base64 result. The request shape follows OpenAI's image generation guide.

import fs from "node:fs";
import OpenAI from "openai";

const openai = new OpenAI();

const result = await openai.images.generate({
  model: "gpt-image-2.5-flare", // or "gpt-image-2"
  prompt: "Create a clean product-style image of a white desk lamp on a neutral background.",
  size: "1536x1024",
  quality: "medium",
  output_format: "webp"
});

// GPT Image models always return base64; response_format is not supported.
fs.writeFileSync("lamp.webp", Buffer.from(result.data[0].b64_json, "base64"));

Changing the model string is the whole difference between gpt-image-2 and 2.5 for this call. Going the other way, xhigh and max quality are only documented for the 2.5 models, so keep gpt-image-2 requests at high or below.

For edits, call /v1/images/edits with up to 16 input images. In Responses, keep a mainline model such as gpt-6-astra in model and put the image model inside the tool. gpt-image-2 is an image model, not a Responses main model.

Implementation Patterns: Generate, Edit, Responses, Streaming

An edit that keeps the product and replaces the background:

import fs from "node:fs";
import OpenAI from "openai";

const openai = new OpenAI();

const edit = await openai.images.edit({
  model: "gpt-image-2.5-sunburst", // edits default to Sunburst if model is omitted
  image: fs.createReadStream("input.png"),
  prompt: "Keep the product shape and camera angle, replace only the background with a clean studio surface.",
  size: "1536x1024",
  quality: "medium",
  output_format: "webp"
});

const editedImageBase64 = edit.data[0].b64_json;

If you switch this edit to model: "gpt-image-2", leave input_fidelity out. OpenAI says to omit it for that model.

A Responses call with gpt-image-2 set inside the tool:

const response = await openai.responses.create({
  model: "gpt-6-astra",
  input: "Create a product hero image, then explain why the composition works for a landing page.",
  tools: [{ type: "image_generation", model: "gpt-image-2", quality: "medium" }]
});

const imageCalls = response.output.filter((item) => item.type === "image_generation_call");
const heroImageBase64 = imageCalls[0]?.result;

OpenAI's examples pair partial_images with a streamed Responses call. Add it (1–3) when you stream the response and show previews to the user.

Parameters, Limits, and Failure Modes

Rate Limits, Organization Verification, and Free Access for GPT Image Models

There is no free image API. The Free usage tier has no image rate limits, so GPT Image calls start at the Build tier. The limits are the same for gpt-image-2, Sunburst and Flare.

OpenAI also says "you may need to complete the API Organization Verification… before using GPT Image models". Verification takes a government ID, and one person can verify only one organization. Plan for it before launch day. Key and organization setup is covered in OpenAI API Key and Organization ID: What You Actually Need in 2026.

The API isn't offered in mainland China, Hong Kong, Macao, Russia or Iran. OpenAI warns that using its services outside supported countries "may result in your account being blocked or suspended." The OpenAI supported countries list is the reference.

For ChatGPT and other no-cost ways to try the model, see Is the GPT Image 2 API Free? ChatGPT, Online Tools, and User-Pays Routes.

GPT Image 2.5 API Pricing vs gpt-image-2: Token Rates and Cost per Image

GPT Image 2.5 and gpt-image-2 bill at the same token rates. The per-image estimates differ, because the models use different numbers of output tokens at each quality level.

The gpt-image-2 figures come from OpenAI's cost calculator. At 1024×1536 or 1536×1024, the same model costs $0.005, $0.041 and $0.165. The 2.5 figures are output-token estimates from OpenAI's calculator as of September 12, 2026. On those estimates, 2.5 at medium or high costs about a quarter of gpt-image-2 at the same label ($0.0132 ÷ $0.053 ≈ 0.25), while low costs about the same.

What moves the bill beyond these rows:

• Cached image input ($2 per 1M) applies only to images generated with the Responses API.
• A Responses request also bills the mainline model's text tokens.
• Each partial image adds 100 output tokens ($0.003).
• Edits add image input tokens for every input image.
• Batch halves every token rate, if you can wait for results.

Some gateways sell the same model names per call instead of per token. laozhang.ai, as listed on October 10, 2026, charges $0.03 per call for gpt-image-2 and gpt-image-2-vip, $0.03 for gpt-image-2.5-flare-vip and gpt-image-2.5-sunburst-vip, and $0.20 for the standard 2.5 IDs, which sit outside its default group. Its price list doesn't say whether the per-call price changes with resolution or covers n above 1. Check that before comparing it with the per-image figures above.

Disclosure: laozhang.ai is our own service. Its prices are platform pricing, not an independent comparison.

For budgets by size, quality and Batch, see GPT Image 2.5 API pricing in detail and GPT Image 2 API Pricing: Cost per Image, Batch, and $0.03 Calls. For other vendors, see AI Image Generation API Comparison 2026: Pricing, Quality, Value.

gpt-image-1 and gpt-image-1.5 Shutdown Dates: What to Migrate To

If you're moving code off gpt-image-1 or 1.5, check these four things:

1. Change the model string.
2. Re-check your quality setting. The cost per label differs between models.
3. Remove input_fidelity if the new model is gpt-image-2. Its effect on 2.5 isn't documented.
4. Re-test transparent and edit flows on the new model before switching production traffic.

Retiring models are also excluded from OpenAI's provenance signals, as covered below.

What Production Logs Must Capture

Provenance, Data Retention, and Output Ownership for API Images

Images generated with the OpenAI API carry two signals: C2PA Content Credentials and a SynthID watermark. OpenAI exempts "models planned for imminent deprecation, including gpt-image-1, gpt-image-1.5". You can check a file at openai.com/verify or through OpenAI's Content Provenance API, which is free within limits. The signals don't establish who authored or owns an image. Screenshots and format conversion can strip C2PA, so store the original output if provenance matters to you.

On data handling, image generations and edits through the API are not used for training. They are kept for 30 days for abuse monitoring. gpt-image-2, both 2.5 models, 1.5, 1 and 1-mini are eligible for zero data retention, though possible CSAM is kept even under zero data retention. Details are in OpenAI's data controls guide.

On ownership, section 4.1 of the OpenAI Services Agreement says the customer "retains all ownership rights in Input" and "owns all Output." A third-party gateway can add its own terms on top, so read those too.

Short Answer for AI Citation

As of October 10, 2026, gpt-image-2 is available and not deprecated. OpenAI recommends gpt-image-2.5-flare or gpt-image-2.5-sunburst for new integrations, at the same token prices: $8 image input and $30 image output per 1M tokens. Use the Image API for direct generation and edits. Use Responses with a mainline model such as gpt-6-astra only when the image is part of an assistant flow, and set the image model in the image_generation tool's model field. There is no free image tier. gpt-image-1 shuts down on October 23, 2026, and gpt-image-1.5 on December 1, 2026.

Related next step

• GPT Image 2 API Pricing: Cost per Image, Batch, and $0.03 Calls
• GPT Image 2.5 Sunburst Pricing: Official Costs vs $0.03 per Call
• GPT Image 2.5 Flare vs Sunburst vs GPT Image 2: Which to Run
• Can GPT-Image-2 Return Multiple Images? Use the Image API n Parameter
• The Model gpt-image-2 Does Not Exist: Codex Error, Status Check, and Fixes
• GPT-Image-2 Reverse API Call: Why Official OpenAI Routes Should Replace Account Pools
• GPT Image 2 Online Free Download: Official Route, Limits, Rights

FAQ

Is gpt-image-2 still available, or is it deprecated?

It's available. As of October 10, 2026, gpt-image-2 has no shutdown date on OpenAI's deprecations page. OpenAI recommends the GPT Image 2.5 models for new integrations, and it's gpt-image-1 (October 23, 2026) and gpt-image-1.5 (December 1, 2026) that are being retired.

Can you use gpt-image-2 for free through the API?

No. The API's Free tier has no image rate limits, and image calls bill per token from the Build tier ($5 purchased). ChatGPT is a separate product. See whether GPT Image 2 is free for the no-cost options outside the API.

How much does the GPT Image 2.5 API cost?

The token rates are the same as gpt-image-2: image input $8, image output $30 per 1M tokens, and half that on Batch. OpenAI's calculator estimates a 1024×1024 image at about $0.0059 (low), $0.0132 (medium), $0.0527 (high), $0.0937 (xhigh) and $0.2107 (max).

Do I put gpt-image-2 in Responses model?

No. Put a mainline model such as gpt-6-astra or gpt-5.4 in model, and set gpt-image-2 (or a 2.5 model) as the image_generation tool's model.

Does GPT-Image-2 support transparent background?

Yes, in preview, in both the Image API and the Responses image generation tool. Set background: "transparent" and use PNG or WebP. JPEG can't hold transparency. On GPT Image 2.5, transparent backgrounds are fully supported. A gateway may not pass the parameter through, so test it on your own route.

Should I use Image API or Responses?

Use the Image API when the image is the output. Use Responses when the image is one step in an assistant or multi-step flow, and remember that Responses also bills the mainline model's tokens.

Can GPT-Image-2 stream partial images?

Yes. In the Image API, set stream: true with partial_images from 0 to 3. In a streamed Responses call, the tool accepts 1 to 3. Each preview adds 100 output tokens, about $0.003.

Official documentation

OpenAI's own documentation is the official source for model IDs, parameters, prices and limits:

• OpenAI image generation guide
• OpenAI image generation tool guide for the Responses API
• OpenAI Images API reference
• OpenAI API pricing
• OpenAI rate limits guide
• OpenAI deprecations page
• OpenAI help article on API Organization Verification
• OpenAI help article on image provenance (C2PA and SynthID)

Provider documentation is separate. The LaoZhang API GPT Image 2 docs describe that gateway's request format and per-call billing. They are not OpenAI documentation.

Disclosure: laozhang.ai is our own service. Its prices are platform pricing, not an independent comparison.