GPT Image 2 API Guide: Pricing, Endpoints, and First Request (May 2026)

GPT Image 2 API access opens the first week of May 2026. That's also the week DALL-E 3 shuts down (May 12, 2026 is the hard cutoff). If you've shipped anything against DALL-E 3 — and most production AI image pipelines have — this is your migration window. This guide walks through the GPT Image 2 API endpoint shape, the pricing that's been confirmed in OpenAI's developer announcements, an authenticated request you can copy-paste, and the migration path off DALL-E that minimizes broken behavior in your existing code.

What the GPT Image 2 API ships with

The GPT Image 2 API exposes the same capabilities ChatGPT users got on April 21: 2K native output, 4K upscale, multi-frame generation up to 8 frames per call, and an optional Thinking mode for complex compositions. The endpoint matches the pattern of the existing OpenAI image APIs — POST to /v1/images/generations with a model parameter set to gpt-image-2. The response shape adds two fields: a frames array (length 1–8 depending on the request) and an upscale_url that's populated when 4K upscaling is requested.

Pricing (confirmed)

Pricing per image lands roughly 20% above GPT Image 1, with multi-frame calls billed per frame. Standard mode at 2K runs ~$0.045 per frame; 4K upscaled output runs ~$0.07 per frame. Thinking mode applies a flat 1.8× multiplier on whichever resolution you generate at. Multilingual prompts are billed identically — there's no language premium.

Your first GPT Image 2 API request

The minimum-viable request is six fields: model, prompt, size, n (number of frames), response_format, and quality. The default response_format is url, which returns time-limited URLs to the generated images. Use b64_json if you need to ingest the bytes directly into a pipeline without a second HTTP fetch.

Single-image cURL example

curl https://api.openai.com/v1/images/generations \ -H "Authorization: Bearer $OPENAI_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "gpt-image-2", "prompt": "A transparent isolated red maple leaf on white background, photoreal, 2K", "size": "2048x2048", "n": 1, "quality": "standard", "response_format": "url" }'

Multi-frame request

For animation source material, set n to a value between 2 and 8 and add a frames_consistency: "high" parameter. The model maintains character identity across the returned frames so you can stitch them directly into a GIF or sprite sheet.

GPT Image 2 vs DALL-E 3: API migration path

If you're migrating off DALL-E 3 before May 12, you have two options. Option one — minimum-change migration — swap the model field from dall-e-3 to gpt-image-2 and you're functional. Output quality jumps but cost goes up roughly 20% per call. Option two — optimized migration — refactor the call to use multi-frame mode where it makes sense, drop external upscaling steps now that 4K is native, and consolidate any post-generation text-overlay code (GPT Image 2's text rendering removes the need to overlay captions externally for most use cases).

Three things that will trip you up

First, image dimensions. DALL-E 3 accepted 1024x1024, 1024x1792, 1792x1024. GPT Image 2 's allowed sizes are 1024x1024, 2048x2048, 2048x1152, 1152x2048. If you have hardcoded size strings in client code, audit them now.

Second, response time. GPT Image 2 in standard mode runs ~25% slower than GPT Image 1 due to the higher native resolution. In Thinking mode it's roughly 2× slower. If you've got synchronous user-facing code paths, push these to async or you'll start tripping timeouts.

Third, content policy returns. The error shape stayed identical, but the policy itself tightened on photoreal celebrity likenesses and explicit branded character generation. Anything that worked on DALL-E 3 in those areas may now return a content policy error. Add error-handling for the policy_violation code if you don't have it already.

Building a transparent-image pipeline on GPT Image 2

Frequently asked questions

The bottom line

GPT Image 2 API access is the inflection point that turns generative imaging from a single-image tool into an animation primitive. Frame-consistent multi-image responses, native 2K output, and the simplest migration path OpenAI has shipped in years. If you're on DALL-E 3 today, your migration deadline is May 12 — and the version you're migrating to is genuinely better, not just newer.