Images and Voice in ChatGPT: Image Generation, Image Inputs and ChatGPT Voice

Version note: This post describes image generation, image inputs and ChatGPT Voice in ChatGPT as of 15 November 2026; the details were checked the same day on the Image generation, Image inputs, ChatGPT Voice and Pricing pages of the ChatGPT documentation (learn.chatgpt.com), the plan list on chatgpt.com/pricing, and the Help Center's release notes and ChatGPT Voice articles (help.openai.com). This post is based on the official documentation. On the voice side there is a retirement on 11 December 2026; if you read this after that date, check the voice settings again.

Guess First

Ayşe was looking for accessibility problems on the Angular basket screen of Bookstore, the online bookstore used as the shared example across this site's series. Instead of the test environment, she took a screenshot of a real customer's basket in the live system; the image also showed the customer's name and delivery address. She told ChatGPT "list the accessibility problems on this screen" and copied one line of the answer into Jira as it was: "The contrast ratio of the campaign label is 3.1:1". The same day Emre asked for an image for the empty basket screen and wrote in his prompt "in the style of a famous children's book illustrator, with their well-known character".

What is your guess? Which of these three steps was wrong: uploading an image with customer data, trusting a contrast ratio read from an image, or asking for a well-known character? The answer is that all three are problems. The documentation on image inputs and image generation describes what to watch out for as much as what they can do.

Image Generation: Where and on Which Plan?

According to the Image generation page of the ChatGPT documentation (https://learn.chatgpt.com/docs/image-generation, 15 November 2026), you can ask ChatGPT to generate an image or edit an existing one: UI assets, banners, backgrounds, illustrations, placeholders. Built-in image generation in Work and Codex uses the gpt-image-2 model and counts toward your general usage limits; the documentation says an image generation uses included limits 3–5x faster on average than a similar turn without one.

On plans, two official pages say different things. The Pricing page on learn.chatgpt.com says, in the context of Work and Codex, that image generation is not available on the Free plan; chatgpt.com/pricing lists "limited image creation" for Free, more images for Go and unlimited, faster image creation for Pro. A reasonable reading: Free has a limited allowance in Chat but none in Work and Codex on the desktop. Availability and limits on the web depend on your plan and workspace settings; check your own account.

According to the Help Center release notes, ChatGPT Images 2.6, released on 4 November 2026, brought two new things: GIF creation, rolling out gradually, and saved reference images. You can save a photo under a name and mention it with @ in later chats without uploading it again. When you open a generated image, Focused view shows a single image and Canvas view shows all images generated in the same chat; with Comment you add notes to one or more images and with Multi-select you send the selected ones for editing in the same chat.

A Good Image Prompt and Targeted Edits

The documentation says a good image prompt is often one to three clear sentences and lists what it should contain: the image's purpose and audience, the main subject and what is happening, the setting, composition and style, size, lighting and colors when they matter, and what the image must not contain. Concrete wording, such as saying where the light comes from instead of asking for "beautiful lighting", gives better results. For Emre's empty basket image, a prompt like this is enough:

Create a simple illustration for the empty basket state of an online
bookstore. Show an open, empty shopping basket next to a small stack of
three closed books on a plain light background. Flat, original style with
two accent colors, generous empty space on the right for a short message.
Square, 1024x1024. No text, no logos, no existing characters or brands.

When editing, the rule is to say clearly what should change and what must stay exactly the same. The documentation suggests changing one element at a time so the composition and other details do not drift. You can also select a specific area of the image and ask for a change to that area only. When you give several reference images, name each one by order and say which one defines the content and which one the style. If the image needs text, put the exact text in quotation marks with the capitalization you want, describe the font and placement, and say that no other text is allowed.

The answer to the second problem in Guess First is here too. The documentation recommends asking for an original design instead of imitating a specific brand, product, artist or artwork, and asks you to confirm that you have permission to use the likeness of a real person. You do not have to credit OpenAI for a generated image, but its use must follow OpenAI's usage policies and your organization's rules.

What Is It Good for as a Developer?

Image generation is a quick start for a UI mockup draft, an empty-state illustration or an image for a presentation. The documentation says it can also help draft explainers, posters, labeled diagrams and timelines, but for dense copy and production-critical typography every word must be reviewed and the asset finished in a design tool when needed.

For architecture diagrams this warning matters even more. The boxes and arrows in a generated image can look convincing while containing a wrong dependency, a wrong protocol or a service that does not exist, and fixing it means generating it again. Producing the real diagram with a tool that draws from text, in other words with code, is safer: its source can be reviewed, versioned and redrawn when things change. The figures in this series were made that way, generated from code as SVG. Use image generation at the idea stage and draw the final version with code.

Image Inputs: Asking with a Screenshot

The Image inputs page (https://learn.chatgpt.com/docs/image-inputs) recommends adding an image when the task depends on visual context: an error screenshot, an interface design, an architecture diagram, an existing asset. On the web you attach, paste or drag the image into the composer. The main rule in the documentation is not to leave the task to the image alone: say what the image shows, point to the area that matters, and state the output and constraints you want. With several images, name each one and explain how they should be compared. Image inputs are for inspecting something; image generation is for creating or editing something.

Everything read from a screenshot is an estimate. It is a good start for reading text, recognizing an error message or spotting layout problems; but information such as a contrast ratio calculated from pixel colors, a hidden label or keyboard order should not go into a ticket before it is checked in the code and the browser. The ratio in Ayşe's Jira ticket was exactly this kind of information.

Bookstore: An Accessibility Review of the Basket Screen

Let us rebuild Ayşe's task. The screenshot is taken in the test environment with sample customer data; the address field contains only a sample city. The prompt says what to look at in the image and what the output should look like:

The attached image is a screenshot of the basket page of our Angular app,
taken in the test environment with sample data. Review it for likely
accessibility problems against WCAG 2.2 AA. For each finding give: the
area of the screen, the likely success criterion, why you suspect it, and
how a developer should verify it in the code or the browser. Mark which
findings you can see directly in the image and which are guesses that
need checking. Do not estimate exact contrast ratios from the image.

The expected answer is a list of findings area by area: a campaign label that looks low in contrast, a delete button shown only as an icon, quantity buttons that look small, and the fact that the image cannot show whether a visible focus indicator exists. Before moving this list into tickets, each line has to be verified according to its type. The tools in the Accessibility (a11y) series post "Accessibility Test Automation: a11y Checks in CI with axe-core, Lighthouse, Playwright and ESLint, and the Limits of Automation" are for this job:

The table says something simple: image inputs speed up the question, not the measurement. The list from a screenshot is not an audit; it is a list of where to look.

ChatGPT Voice

There are two separate voice experiences, and the documentation describes them on separate pages. The first is the voice conversation in Chat. According to the Help Center article "ChatGPT Voice" (https://help.openai.com/en/articles/20001274), Live voice runs on the GPT-Live model; because it can listen and speak at the same time, interrupting and taking turns feel more natural. Free users get limited access to the smaller GPT-Live-1.1 Quip model, and paid plans use the full GPT-Live-1.1 for longer. According to "Retiring Advanced and Standard Voice in ChatGPT" (https://help.openai.com/en/articles/20001638), Advanced and Standard voice retire on 11 December 2026; after that date they can no longer be selected in Settings > Voice, and people who had selected them are moved to Live voice. According to the article, Live does not support video or screen sharing yet; those features stay on Advanced voice for now. How they will be offered after the retirement is not clear in the article, so check again on that date.

The second is ChatGPT Voice in the desktop app. According to the ChatGPT documentation (https://learn.chatgpt.com/docs/features/voice), it uses GPT-Live to let you work by talking in Chat, Work and Codex: start a task by voice, ask about the status of running tasks or change direction. It is available on the Plus, Pro, Business, Edu and Enterprise plans. Voice follows the same permissions as the tasks it directs, and only one voice chat can be active in the desktop app at a time. According to the Pricing page, voice on the desktop uses your Codex usage budget at $0.05 per minute, and 1.25 credits per minute in credit-based Business, Edu and Enterprise workspaces; the model handling the task is billed separately.

On macOS, with Screen context turned on under Settings > Voice, saying "Take a look at this" makes ChatGPT take an appshot of the frontmost window. The documentation gives a clear warning here: an appshot can include the window's image and accessible text, including content outside the visible scroll area; do not share windows that contain sensitive information. Your organization can turn this capability off.

In Production

Clean a screenshot before uploading it. If a customer's name, address, e-mail, order number or payment details are visible, either take it from the test environment or blur them. We will look at how personal data reaches ChatGPT and which settings apply in detail in the tenth post.

Know the setting for voice data. According to the Help Center, if "Improve the model for everyone" is on, transcripts of voice conversations may be used for training; audio and video recordings are used only if you turn that on separately under Data Controls.

Count it against your limit. Image generation and desktop voice draw from the same pool as Work and Codex. Instead of asking for ten variants for a presentation, refining one prompt and making two or three attempts is both cheaper and easier to control.

Common Pitfalls

Leaving customer data on screen. A screenshot carries as much data as text you copy, and often more than you notice.

Trusting a number read from an image. A contrast ratio, a pixel size or a figure in a table may have been estimated from the image; measure it at the source.

Asking for a brand or a well-known character. Ask for an original design; an image that imitates someone else's work should not end up on a product screen.

Treating a generated diagram as documentation. A wrong arrow in an image is easy to miss. Keep the architecture in a diagram drawn with code.

Forgetting the retirement date. After 11 December 2026, Advanced and Standard voice cannot be selected; if your team has habits that depend on these modes, try Live voice now.

A Small Experiment

★☆☆ Generate an image with the empty basket prompt from this post. Then ask it to change only the color of the stack of books and check that everything else is kept.

★★☆ Run the accessibility prompt from this post with a test-environment screenshot of a screen from your own project. Split every finding into "seen in the image" and "needs checking" as in the table, and verify at least three with a tool.

★★★ Write an image rule for your team: which jobs use image generation, which diagrams are drawn with code, what is cleaned before a screenshot is uploaded, and where generated images are kept.

Screenshots, voice recordings and uploaded files are all data that goes to ChatGPT. In the tenth post we will look at the model training setting, Temporary Chat, memory, the defaults of personal and business plans and KVKK practice, with a rule list for Bookstore's support team.