What is a multimodal model? Images in the prompt

By

A multimodal model takes more than text, such as a screenshot, in the same request as the prompt. How Claude Code, Codex, Cursor and Grok accept an image.

A multimodal model can work with more than text. A screenshot, a diagram, or a photo can sit in the same request as the prompt.

Images are the case the coding tools document. Audio and video show up on some model cards as a separate limit, not as a file you drop into the chat.

GPT-6.1 Terra is one vision model in the Codex lineup. Moondream on Workers AI is a small vision model on Cloudflare. This entry is how the four coding tools attach an image.

How a multimodal model works

The request carries the picture and the words together. The model reads both, then answers or edits the code. You still have to say what you want. The image is evidence, not the task.

An image joins the promptYou attach a screenshot and write a prompt. The model reads the image and the text together, then answers or edits the code.ScreenshotModelPromptAnswer or a codeedit

A file read of Button.tsx is exact. A screenshot of that button is what the screen showed, including the crop and the blur.

Multimodal models in each tool

Claude Code Codex Cursor Grok
Attach an image Drag, paste, or a file path Paste, or --image / -i Read includes image files for vision-capable models API input_image. Not documented in the Grok Build TUI
Generate an image Not on the workflows page A different feature from image input A separate Image generation tool /imagine generates. It does not read a file you already have

Claude Code

The common workflows page, under “Work with images”, gives three ways in. Drag and drop into the Claude Code window. Copy an image and paste with Ctrl+V, or Alt+V on Windows and WSL. Or give Claude a path.

You can keep several images in one conversation. The page points at screenshots of errors, UI designs, and diagrams. When Claude refers to [Image #1], Cmd+Click on a Mac or Ctrl+Click opens it.

Computer use is a different page. It is not the same as pasting a PNG. See Work with images.

OpenAI Codex

Paste into the interactive composer, or pass a file:

codex -i screenshot.png "Explain this error and suggest the smallest fix"

Several files are a comma-separated list, or repeated --image flags:

codex --image before.png,after.png "Compare these states and list the regressions"

The image inputs page accepts common formats, including PNG and JPEG. In the IDE extension, drag while holding Shift.

The Codex models page calls GPT-6.1 Terra a fast choice for writing, visual understanding, and computer use. It also says Terra needs Codex CLI 0.159.0 or later: codex -m gpt-6.1-terra. The API model card for gpt-6.1-terra, checked in November 2026, lists input as text and image, and output as text. Audio and video are unsupported on that card.

Image input and image generation are different requests. See Image inputs and How to generate images with AI.

Cursor

The agent page says Read also supports image files (.png, .jpg, .gif, .webp, .svg) and includes them for vision-capable models. The page does not name which models those are.

The Browser tool takes screenshots. Image generation is a separate tool. Generated images save to the project’s assets/ folder by default. See the agent overview.

Grok

The image understanding page says some models accept images. You send input_image with image_url set to a base64 data URL or a public URL, plus input_text. The page’s example model is grok-4.7.

Limits on that page, and again on the models index: 20MiB, no maximum image count, and jpg/jpeg or png. Image and text can be in any order. The page warns that storing request and response history on the server can make an image request fail.

The Grok 4.7 card says the modalities are text and image in, text out. The models index lists a separate Voice API, and separate image and video generation models. Those are not this attachment.

Grok Build’s command list has /imagine <prompt> to generate an image. The CLI reference this pass checked has no image or screenshot flag for the TUI. Attaching a picture there is not documented. See Image understanding.

Details checked against each vendor’s documentation in November 2026.

A real example

The Codex command above is the one to copy. screenshot.png is a file on your machine, the same way the docs use it. The prompt names the job: explain the error, and keep the fix small.

A path in Claude Code works the same way. Point at the file, then say what to look for. “What’s causing it?” on a screenshot of a stack trace is the workflows page’s own example.

When to use a multimodal model

Use it when the bug is on screen. A layout that wraps, a dialog with the wrong label, a diagram in the repo, a before and after of a change.

When the failure is a log line or a test name, paste the text or let the agent read the file. Pixels are the worse copy of a file you already have.

Pitfalls

A screenshot of source code throws away the file. The model sees the crop. Read the file when you have it.

Terra’s API card does not accept audio or video. A screen recording is not an input that model documents.

Grok Build’s /imagine creates a picture. It does not explain one you already saved. Cursor’s image generation tool is the same split.

The Grok image API can fail if the request asks the server to store history. That warning is on the understanding page.

VM Peek is the app here for watching a Mac test VM while a coding agent drives it. That is a screen the agent can look at. It is not itself a vision model.

  • Frontier model: A frontier model is one of the most capable models right now, and the label moves. How Claude Code, Codex, Cursor and Grok name the model they point you at.
  • Training data: Training data is what a model learned from, and it stops at a cutoff. Dates on the Grok 4.7, GPT-6.1 Terra and Claude cards, and how an agent reads your repo.
  • Coding agent: A coding agent uses a model and tools to read a repo, edit files and run commands in a loop. How Claude Code, Codex, Cursor and Grok Build each do that job.

All AI glossary entries →

Want me to talk about your product? You can sponsor this site.