If you want to create images with ChatGPT Pro, the usage pool you draw from depends on whether you create them in regular "Chat" or in "Codex / Work". The ChatGPT pricing page lists image creation as "unlimited" for regular Pro chats. Image generation in Codex draws from the Codex / Work pool, which is shared with coding tasks. Claude has no built-in ability to draw photos or illustrations, but Claude Code can draw diagrams with code or call external image generation APIs.
What this article covers
- ChatGPT Pro's "unlimited image generation" applies to regular Chat
- Image generation in Codex uses tokens and reduces your work usage
- ChatGPT and Codex images made from the same request
- Can Claude and Claude Code not create images?
- Tips for making usable images: from before generation to selection
- By use case: where is it best to create?
- Frequently asked questions
ChatGPT Pro's "unlimited image generation" applies to regular Chat
OpenAI's ChatGPT pricing page lists "Unlimited and faster image creation" as a Pro benefit. However, the same page carries a note that it is subject to abuse-prevention limits, so limits may apply at times such as during peak demand.
In regular ChatGPT, you either ask for "create an image" within a conversation or start from the image creation screen. As described in OpenAI's image feature help, you can also attach an existing image and ask for changes. If you only need an image, this entry point is the easiest to understand. If you plan to use the image in another tool, save the full-size image file.
The easy-to-confuse pair is "Chat" and "Work" in ChatGPT. Even in the same ChatGPT interface, Work shares its usage pool with Codex. The Work / Codex pricing explanation states that usage pools and credits are shared between the two, so Pro's "unlimited" does not apply to Work or Codex.
Image generation in Codex uses tokens and reduces your work usage
You can generate and edit images in Codex too, but it is handled separately from the image limits on regular ChatGPT chats. According to the OpenAI Codex help article, the image generation limits and reset times on the ChatGPT side do not apply to Codex. In Codex, image inputs and outputs are processed as tokens, and when you use it through a subscription, they are first counted against your plan's general usage pool.
Official average guideline: compared with a similar turn without image generation, image generation uses your usage pool about 3 to 5 times faster. The amount varies with image quality and size, so you cannot convert it to "one image = how many messages."
If you are a Pro user logged in with a ChatGPT account, generating an image does not separately bill you at API rates each time. First the image uses the allowance included in your plan. Once that is used up, it draws from additional credits if you have any, and otherwise you wait for the allowance to recover. On the other hand, if you set up Codex with an API key, API charges apply. You can check your current balance and reset time on the Codex usage screen.
If you have Codex create images for your site and then also have it check the images and fix the code, you can finish the whole job in one workflow. If you only need to repeatedly make image candidates, you can instead generate them in regular Chat, save the full-size files, and pass only the images you choose to Codex. The key is to split the work deliberately according to which pool you are using.
Before you start, check not just the screen names but also your current mode and authentication method. Is your ChatGPT conversation a regular Chat or Work? Are you logged in to Codex with a ChatGPT account or an API key? If you mix up these two points, explanations of remaining usage and charges will not match, even if you say "I made the image with Pro" in both cases.
ChatGPT and Codex images made from the same request
Let's look at actual examples. On September 30, 2026, we entered the same English prompt once each for this site. The ChatGPT side used a regular Chat on a Pro account, and the Codex side used its built-in image generation. We did not attach reference images, make edits after generation, or select from multiple candidates.
From public documentation, we can tell that OpenAI's Responses API lets you choose a setup in which the conversation model calls an image generation tool and a GPT Image model creates the image. The conversation-side model may automatically rewrite the instructions for the image. OpenAI has not disclosed whether regular ChatGPT Chat and Codex use the same image model and the same internal instructions.
View the full shared prompt
One horizontal concept art image. An observatory on a sea cliff with a giant celestial sphere, teal and amber light, and small human figures placed to convey scale. The prompt specifies no text or logos.
Use case: stylized-concept. Asset type: editorial sample image for a Japanese article demonstrating AI image generation. Create ONE standalone horizontal 16:9 image, without any typography. Scene: a vast glass-and-brass observatory built into a sheer black volcanic cliff above a storm-lit ocean at twilight. An immense intricate armillary sphere hangs over the sea, its concentric rings catching warm amber light; narrow bridges and a few tiny human silhouettes reveal the colossal scale. Below, turquoise bioluminescent waves strike the rock; above, layered storm clouds break into shafts of light and faint stars. Style: exceptionally polished cinematic concept art with believable architectural detail, atmospheric depth, coherent perspective, restrained deep teal and amber palette, tactile materials, sharp focal point and rich detail that survives reduction to an article card. Composition: one wide coherent scene, focal sphere slightly right of center, clean readable silhouettes. No words, letters, signage, logos, borders, collage, or watermark.
Can you tell from the image file that it was made with AI?
OpenAI explains that generated images from supported ChatGPT and Codex features include C2PA provenance metadata and an invisible SynthID watermark. Whether they are included depends on the product, model, how the file was saved, and when it was created. C2PA is information attached to the file, and it can be lost through conversion or re-saving. SynthID is a separate signal embedded in the image content. As described in OpenAI's provenance guide, failing to detect either one does not mean the image is not AI-made.
caBXcaBX is the area that the C2PA specification defines for manifest storage in PNG. The presence of that area and a valid signature are two different things.We also provide the original PNG from ChatGPT and the original PNG from Codex before conversion. The OpenAI verification tool can check the C2PA and SynthID of the corresponding files, but it means sending the image to that service. What it reveals is the provenance signal only. It cannot tell you the prompt, the user, or which screen the image was made on.
If you want to check only SynthID, you can also use the SynthID Detector, which Google made available to anyone on October 7, 2026 (it supports images, video, and audio, and is offered worldwide in English; checked October 18, 2026). It can identify only content generated by companies that use SynthID, such as Google and OpenAI. It is not a tool for identifying all AI-made content.
Can Claude and Claude Code not create images?
If you mean creating photos or illustrations with an image generation model, current Claude models do not have that capability. Anthropic's official help article explains that Claude does not create photos or illustrations the way an image generation tool does, and the API vision documentation also separates the ability to understand and analyze images from the ability to generate and edit them.
For diagrams and comparison charts, 01 is a good choice because it makes it easier to place text and shapes accurately. If you need photos or illustrations, use 03. For example, you could set up Claude Code calling OpenAI's GPT Image API or Google's Gemini image API. Claude Code handles organizing the instructions and saving files, while the connected service creates the image itself. You need to prepare the connections to external APIs and MCP tools, as well as the authentication credentials for the image API, yourself.
Even when you go through Claude Code, image API fees still apply, and the Claude side also uses your usage pool or API tokens. With a Claude Pro, Max, or Ultra plan, the Claude side may fall within the plan's usage pool rather than being billed separately (see the Claude Code pricing documentation). The OpenAI image API also counts image generation and the conversation model separately. As an example, the Google Gemini API pricing page lists, as of September 30, 2026, 0.067 dollars per 1K image output for Gemini 3.1 Flash Image on the paid Standard tier (inputs and other items are billed separately).
On the other hand, built-in generation in Codex under a ChatGPT account and direct generation in the Gemini app each draw from that product's usage pool. If you already have allowance within your plan and only need to create image candidates, generating directly in that product saves you from setting up external APIs or managing two separate usage pools. Going through Claude Code makes the most sense when you want to automate image generation, integration into code, and verification as one continuous process.
Tips for making usable images: from before generation to selection
Whichever entry point you use, following these steps in order makes it less likely that you will mass-produce unusable images or mix up candidates.
1. Decide the size it will be used at before the prompt
An image that looks fine at full size may lose its main subject when shrunk to a small card in a list. If it will be displayed small, emphasize the features that help people tell it apart, such as the face, hairstyle, and clothing, rather than fine details of the full body. Composition instructions such as "square," "from the chest up," "make the face larger," and "prioritize the person over the background" are effective. For a wide hero image, also decide in advance where the space for text overlays is and which parts will be cropped.
2. Separate reference material from what is newly created
If you want to draw an existing character or product, don't leave it to the name alone. Provide the rights holder's official introduction page and the image URL. Take only the visual features from there, and decide the pose, composition, background, light, and line work anew. Leave out third-party images that have been reposted, and it is also effective to specify that generation should stop if the reference material cannot be opened. If you keep the URLs of the references, you can later trace what the drawing was based on.
That said, referencing official images does not make them legally safe to use. Check the terms for the reference material separately from the rights and similarity of the finished work, and have a person confirm that the result is not just the official image with a different art style applied.
3. Generate a few and judge by both original size and display size
It is less likely that you will mix up images and subjects if you first check the art style, how well features are reproduced, and separation from the background with a small number of candidates, and only then increase the count. When requesting multiple images, specify "one image per subject" and "as separate image files." Without this, multiple subjects may end up merged into a single collage.
Save the full-size files of the generated results, not screenshots of the screen. At full size, check the hands, clothing, and any stray text or logos. At the actual display size, check whether the main subject can be told apart. Keeping a one-line note for each of the file name, save location, reference URL, and reason for accepting or rejecting it means you won't hesitate when handing it to another tool or person.
If your first candidate has a good background but the face is too small to recognize on the card, the next request should be a narrow change, such as "keep the background, move the face into the upper half of the frame," rather than changing the whole art style. If the face is readable but the clothing features have disappeared, widen the range from the chest up a little. Changing only one or two things makes it clear which instruction worked. Once it's finished, review it at the sizes it will actually be used at, such as the list card, the article body, and the social media share preview.
Even if you instruct "no text," text can still sneak in, such as a banner with a name added below the image. In some cases you can crop out the banner and use the rest, but if you select the image by looking only at the subject, you will miss it. Check the edges of the image at full size.
"Use: small list card. Square format, with an exploration robot drawn large from the chest up. A composition where the features of the head and face remain recognizable even when shrunk. The background should have light and atmosphere suggesting a night observatory, but should not stand out more than the main subject. No text, logos, or watermarks. First output one image as an independent image file, and do not mass-produce until I confirm the selection."
4. When passing to another tool, hand over the image and its basis together
When you have Codex or Claude Code incorporate an image made in ChatGPT, if you only hand over the file, it becomes unclear which version was the chosen one. Attach a short note with the file name of the chosen image, the page where it will be used, its purpose, the URL of the reference material, and the features that should remain when shrunk. For example: "Chosen: portrait-final.png / Use: square list card / Reference: official settings page / Rejected: portrait-02.png (face unreadable when small)."
With this note, you can prevent someone from importing the wrong candidate or recreating the same image. If you share reference images, also confirm that the material may be redistributed, and before publishing, check that the captions and alt text actually match the picture.
By use case: where is it best to create?
With Pro, regular ChatGPT Chat is the first choice. Save the outputs and pass only the candidates you need to your production environment.
In Codex, you can generate, save, incorporate, and check the screen in one continuous flow. Check your shared pool balance first.
Claude Code's HTML and SVG, or Claude Design, are suitable. If you need photos, connect an external image model.
If you want to compare the art styles and prices of image generation AI tools themselves, see Image Generation Tool Comparison. If you want to start from the basics of prompting, see Introduction to Image Generation, and for the flow of design production, see the explanation of Claude Design. The decision criteria in this article are the entry point, the usage pool, and the type of thing you want to make, rather than a ranking of tools.
Summary
In regular ChatGPT Pro Chat, image creation is listed as "unlimited" on the pricing page (though abuse-prevention limits apply). Image generation in Codex / Work uses the shared work pool and consumes more than a conversation of similar length without images. Claude is not a model that draws photos or illustrations directly, but you can create diagrams and designs with Claude Code and connect it to external image APIs. Once you've chosen your entry point, specify the use and display size first, and then decide what to accept based on both the full-size file and the actual display size.
Frequently asked questions
If I have ChatGPT Pro, is image generation in Codex also unlimited?
No. The Pro "unlimited image creation" on the OpenAI pricing page describes regular ChatGPT Chat. Image generation in Codex / Work uses the shared general usage pool. The image limits of regular Chat also do not apply to Codex as they are.
How many tokens does generating one image in Codex use?
It is not fixed. It varies with the instructions, whether reference images are included, and the size and quality of the output image. The "average of 3 to 5 times" that OpenAI gives is a comparison with a similar turn without an image, and it does not represent the price of one image or how many remain.
If Claude Code can output a PNG, can't Claude also generate images?
It depends on how it is output. If Claude Code creates an image file from a PNG drawn with SVG or a program, you can say that Claude Code made the image file. However, the Claude model itself has not become an image generation model for photos or illustrations. If the image was made with an external API, the model that generated it is the one it connects to.

