GPT Image API Pricing: When GPT Image 2.5 Costs Matter

That bill is separate from ChatGPT. A ChatGPT subscription lets a person create images by hand in ChatGPT Images 2.5, while the API lets software generate and edit images at volume and is billed on an OpenAI API organization. The two meet only on ChatGPT Pro and Business: their monthly API credits, claimed into an API organization the subscription owner also owns, can pay for Images API usage, and Plus, Go and Free include none. Token rates were read on OpenAI's API pricing page on October 29, 2026, the per-image costs and the gpt-image-1 shutdown on the image generation guide and deprecations page the same day. OpenAI API pricing, Image generation guide, GPT Image 2.5 Sunburst, Deprecations.

Sunburst and Flare replace the GPT Image 1 models at GPT Image 2's rates

OpenAI released GPT Image 2.5 Sunburst and Flare on September 8, 2026. Both share the same prices but target different jobs. Sunburst is OpenAI's most capable image model, built for workflows where editing precision matters most; it gives finer control across edits at the cost of longer generation time. Flare is positioned as fast, high-quality everyday image generation and the default choice for most applications, with up to 50% lower latency than GPT Image 2. Both add new xhigh and max quality settings above the high ceiling of earlier models.

GPT Image 2 (gpt-image-2) is still available at the same token rates, but OpenAI's guide groups it with earlier models and recommends GPT Image 2.5 for new integrations. Older models are retiring: gpt-image-1 shut down on October 23, 2026, and gpt-image-1-mini, gpt-image-1.5 and the deprecated chatgpt-image-latest alias follow on December 1, 2026, with GPT Image 2.5 Sunburst or Flare named as replacements.

Image output at $30 per million tokens is most of the bill

Cached input discounts apply only when you generate images through the image generation tool in the Responses API, not to direct Images API requests such as /v1/images/edits. Output tokens dominate the bill, so quality and size settings matter far more than prompt length. A long, detailed prompt of 500 tokens costs only $0.0025 in text input, less than half of a single low-quality image, so there is little reason to shorten prompts to save money. Spend your optimization effort on quality settings, attempts and the Batch API instead.

Cost per Image by Quality

OpenAI's calculator converts output tokens into a per-image cost and gives one estimate for both GPT Image 2.5 models, though OpenAI says the two can use different token counts at the same quality setting. At 1024×1024 it shows:

These figures cover output only. Add text input for prompts and image input for references or edits; for most generation jobs these are small, but editing workflows that send large reference images can add noticeably to the total. Size moves the price both ways: a 1536×1024 image costs less than a square one at every quality ($0.04116 against $0.05268 at high), while 2048×2048 costs about twice as much ($0.10704 at high). Use the calculator in OpenAI's image generation guide for other dimensions.

What Actually Drives the Bill

Quality. The jump from medium to high is about four times, and from high to max another four times. Most product and marketing images do not need max; use it for hero assets and keep drafts at low or medium.

Retries and variants. If you generate four variants to pick one, your cost per accepted image is four times the per-image price. Track accepted images, not generated images.

Edits. Editing sends the source image as input and generates a new output, so each edit costs at least one more output image plus image-input tokens.

Model choice. Sunburst and Flare cost the same per token, but Sunburst's finer editing control may reduce retries on precise edits, while Flare's lower latency suits interactive features.

Batch. If images do not need to be generated in real time, the Batch API halves output cost. Overnight catalogue generation is a natural fit.

Sunburst or Flare?

Because both models cost the same per token, the choice is about output and speed rather than price. Start with Flare for most generation: product shots, social images, illustrations and anything a user waits for in your interface. Its lower latency also means fewer timeouts and a smoother experience in interactive features. Switch to Sunburst when the job is editing, such as changing one element while keeping the rest of an image intact, matching a brand layout precisely, or applying a sequence of changes to the same picture. Sunburst's finer control can reduce the number of attempts needed for an acceptable edit, which lowers your real cost even though its per-token rate is identical.

Many teams use both: Flare for first drafts and bulk generation, then Sunburst for the final round of edits on the images that will be published. Log which model produced each accepted image so you can see which one actually costs less per accepted result in your workflow.

Estimating a Monthly Budget

A practical estimate multiplies four numbers: accepted images per month, attempts per accepted image, cost per attempt at your chosen quality, and an input allowance. Suppose an e-commerce team needs 2,000 accepted product images a month at high quality, averaging two attempts each. Output alone is 4,000 × $0.05268 ≈ $211. If the work can run in Batch, output falls to about $105. Adding reference images and prompts might add a few dollars to tens of dollars, depending on image sizes.

A marketing team producing 200 hero images at max quality with three attempts each spends 600 × $0.21072 ≈ $126 on output, while drafting the same concepts at low quality first, 600 × $0.00588 ≈ $3.50, and then rendering only the 200 winners at max, about $42, cuts the total to roughly $46. A draft-then-final workflow is the single most effective cost control.

API Versus ChatGPT Images

ChatGPT Images 2.5 is available to ChatGPT users on every tier, as well as ChatGPT Work and Codex users, on desktop, mobile and web. That makes ChatGPT the cheaper route for occasional, human-driven image creation. The API makes sense once generation is automated, repeated or embedded in software.

When the API Is Overkill

The API adds engineering, monitoring, storage and review work. For a handful of images a week, a person using ChatGPT, or a specialist tool such as Adobe Firefly, Midjourney or Recraft, is usually cheaper once you count staff time. The API pays off when you need many images on a schedule, consistent settings across a pipeline, or images generated in response to user actions. If you are comparing vendors, Google's Nano Banana models are another image API; see the Nano Banana profile.

How It Compares With Other Image APIs

OpenAI's per-image costs sit in a similar range to Google's Gemini image models. Nano Banana 2 costs about $0.067 per 1K image and Nano Banana 2 Lite about $0.034, while Nano Banana Pro costs $0.134 at 1K or 2K. GPT Image 2.5 at medium quality, about $0.013, is cheaper than all three, and at high quality, about $0.053, it falls between Nano Banana 2 Lite and Nano Banana 2. Recraft's V4.1 API starts at $0.007 per image for its Flash model. Because quality settings and output styles differ, compare accepted images from each provider on your own prompts rather than relying on list prices alone. Google's newer Nano Banana 2.1 costs $0.0336 per 1K image; Nano Banana 2.1 vs GPT Image 2.5 sets it against each GPT Image 2.5 quality setting.

Migrating From Older Models

Code that still calls gpt-image-1 stopped working when OpenAI shut it down on October 23, 2026, so move it to a GPT Image 2.5 model now; move gpt-image-1-mini and gpt-image-1.5 before December 1, 2026. Test Flare first for general generation and Sunburst for editing-heavy workflows, and compare cost per accepted image rather than per attempt, because the new quality settings can change how many retries you need. Note that GPT Image 1.5 had its own per-image prices, so budgets built on it need recalculating with the GPT Image 2.5 calculator.