What is GPT Image? Names, costs and limits of OpenAI's image model

I used GPT Image 2.5 inside Higgsfield to make my own LinkedIn headshot, so this model is not new to me. That is a different job from explaining it, which is what this page does. If you want the headshot steps, read How to Make a LinkedIn and CV Headshot with Higgsfield. This page is about the model: what it is, what it costs and where it fails.

This page may contain affiliate links from which Blur Studio earns commissions.

The names, explained

OpenAI is the company behind ChatGPT, a chat assistant. It also makes image models. People use several names for the same family, which is why searching is messy.

Name you see	What it usually means
GPT Image	The family name for OpenAI's image models
GPT Image 2 and 2.5	Versions in that family. On 4 November 2026 OpenAI's pricing page and image guide listed only the 2.5 models for new work and called GPT Image 2 an earlier model
gpt-image-2.5-sunburst and gpt-image-2.5-flare	API model names. The docs describe sunburst as the one for editing precision and flare as the fast, high-quality everyday option
ChatGPT Images	The image feature inside ChatGPT. Check which model version your account actually shows
gpt-image-1	An earlier model. It was not on the pricing page, so do not build on it without checking OpenAI's deprecations page
DALL-E	OpenAI's older image line. OpenAI changes its retirement schedule, so check before relying on it

My rule: when a blog says "GPT Image 2", ask which model name it means and what date it wrote the page. Names change monthly. I also could not confirm launch dates from an OpenAI page, so I do not print any.

What it is good at

From OpenAI's docs and the way the model is used, three strengths matter to a business:

• Following detailed instructions. A long, specific prompt (a written instruction) tends to be followed closely.
• Readable text, more than before. OpenAI says text rendering is significantly improved. See the limits below, and the text-on-images guide.
• Editing from reference images. You can give it a real photo and ask for a change. The docs describe mask edits too, where you mark an area, though the guidance is prompt-guided.

Sizes and quality settings matter for cost. OpenAI's docs list quality options of low, medium, high, xhigh, max and auto, recommended sizes of 1024 by 1024, 1536 by 1024 and 1024 by 1536, and custom sizes where both sides are multiples of 16, the ratio is between 1:3 and 3:1, and neither side is over 3840 pixels. It also lists a transparent background option (save as PNG or WebP).

What OpenAI itself says it still gets wrong

• Precise text placement and clarity. Text has improved but can still struggle.
• Recurring characters and brands. The docs say consistency across images is difficult. For a mascot that must look the same in every post, a single prompt is not enough. Use a reference-sheet method, covered in Keep a Brand Mascot the Same in Every Post.
• Layout control. Placement is not pixel exact.
• Mask edits. They are guided by your prompt, and the edge of the changed area may not follow your mask exactly. Always compare the edges.
• Time. Complex prompts can take up to two minutes.

One third-party site said transparent backgrounds are not supported. OpenAI's own guide says they are for both 2.5 models, so trust the guide, and still test your cut-out edges.

What it costs through OpenAI's API

An API is a way for software to call the model directly. Pricing there is by tokens, small chunks of data. On 4 November 2026 OpenAI's pricing page listed both GPT Image 2.5 models (sunburst and flare) at $8 per million image input tokens ($2 when cached), $30 per million image output tokens, and $5 per million text input tokens ($1.25 cached). The image guide says both models use the same rates and that Responses API calls also add the cost of the chat model that calls the image tool.

The key point, which OpenAI states itself, is that equal token rates do not mean equal image costs, because consumption differs by quality setting and size. The guide has a calculator. For a square 1024 by 1024 image at low quality on GPT Image 2.5 it showed 196 output tokens and an estimated image output cost of $0.00588. Costs for medium, high, xhigh and max quality depend on size, so use the calculator for your size and quality.

Step	Worked example (low quality, square)
Output tokens from the calculator	196
Cost = tokens divided by 1,000,000, times $30	196 / 1,000,000 x 30 = $0.00588
Made-up edit: reference photo counted as 1,000 input tokens, times $8 per million	adds $0.008
Total for that made-up edit	about $0.014

The edit line is invented to show the method, not a published token count. Use the usage numbers OpenAI reports back after each call. The practical rule: draft at low quality, and use higher quality only for finals.

Then divide by your pass rate. If one in three images is a keeper, the cost per approved image is three times the cost of one try. This is the number that matters. Example with the low-quality figure: 3 x $0.00588 = $0.01764 per keeper, before any higher-quality final.

How to access it

• ChatGPT. Easiest, no code. Free accounts have stricter image limits than paid ones, but the exact counts change, so check OpenAI's help page for your plan.
• OpenAI API. Token pricing, as above. Best if a developer is building it into your own tool.
• A platform that hosts it. Higgsfield lists GPT Image 2.5 among its image models. There it is priced in credits, not tokens, and the cost shows on the Generate button before you press it.

GPT Image 2.5 inside Higgsfield

I make images with GPT Image 2.5 inside Higgsfield's Images tool because it puts several image models next to each other. Higgsfield is a web platform that sells credits and lets you choose between models from different companies, including Nano Banana Pro (a Google image model) and Seedream (an image model from ByteDance, the company behind TikTok). If your question is "is GPT Image better than Nano Banana Pro for my product photo", the fastest way to find out is to run the same prompt on both on one screen and count the retries. My comparison of other options is in Higgsfield's image models compared and Higgsfield vs Gemini Image and Nano Banana. For the marketing side, see Higgsfield vs ChatGPT Images.

What to know before you pay: on 4 November 2026 the pricing page showed GPT Image 2.5 in a time-limited unlimited block on the Plus and Ultra plans, but promotions change often, so check the pricing page rather than this one. Plan prices and credit costs are on the pricing explainer. The cost of a GPT Image 2.5 run is shown on the button. Anything Higgsfield may add to a file, such as provenance marks, is described in its terms as something it "may embed", with no warranty, so do not remove marks and do not assume they exist.

Who should skip Higgsfield for GPT Image

• You already pay for ChatGPT and make a few images a month. Use the image feature you already have.
• A developer is building it into your product. Use OpenAI's API directly, where you control the quality setting and see the token usage.
• You need a written commercial and data position from the model's maker. Go to OpenAI's terms. Higgsfield's own terms say it may train on content except under an Enterprise agreement.
• You need one model only. A hosting platform pays off when you compare models.

Five jobs it suits

1. Ad and social images with a short headline. Check every letter.
2. Product background swaps. Upload the real product photo and change the background, instead of generating the product from nothing. Logos and labels stay safer that way. See changing one item in an approved image.
3. Diagrams and simple infographics. Good for rough concepts, not for final data.
4. Headshots from reference photos. Check that the result still looks like you.
5. Edits to real photos. Remove an object, change a colour, then compare edges.

Skip it, or add a manual step, for pixel-exact layouts and long-running character series. Transparent backgrounds are listed as supported, but check the edges of any cut-out.

A 5-prompt test against another model

Use the same five prompts on GPT Image 2.5 and on Nano Banana Pro: a headline poster, a product background swap, a menu line, a simple diagram and a portrait from reference photos. For each, score text accuracy, edit fidelity and number of retries. Record the credits or dollars spent. The winner is the model with the lowest cost per approved image on your work, not the one a blog crowns.

Labels, rights and checks before client use

• Provenance labels. Third-party blogs say OpenAI adds C2PA metadata (a standard that records how a file was made) to outputs. OpenAI's own wording was not confirmed, so do not promise a client that a file carries or lacks a mark.
• Commercial use. OpenAI's terms for outputs were not confirmed here. Read them for the route you use.
• Disclosure. Some platforms and laws require you to say when an image is AI-made. Check local law and the platform's current rules. This is information, not legal advice.
• Real people. Never use someone's face without their written permission.

How I'd approach this

Pick one real job, run the 5-prompt test, and write down cost per approved image. Draft at low quality, finish at high, and edit a real photo whenever you can rather than inventing the product. Name the model and version on every saved file so a client question in six months has an answer.

If you make images like this every week, the useful build is a shared library with the prompt, model, version and approval status beside each file. Blur Studio builds that kind of approval and asset workflow around these tools. For a wider choice of image tools, see Best AI Image Generators for Business Use.