Can Claude generate images? What Anthropic says, what we got when we asked, and the routes that make a real photo

No. Claude reads images but does not draw pictures, and Anthropic says so in its Help Center, its developer documentation and its model specifications. That is the short answer, and it is the one most pages ranking for this question stop at, or contradict in their headline to sell you a connector.

The longer answer matters more if you came here because you need a picture for a post. Ask Claude for one today and you do get something back, and what you get depends on where you ask. So we gave Claude the brief a coffee shop owner would actually type, ran it four ways on 12 October 2026, and show every result below: what Claude made by itself, what it refused, what Claude Design produced, and what happened when Claude Code was handed a key to Google's image model.

Can Claude generate images, according to Anthropic?

Anthropic answers the question in its Help Center under exactly that title, in a three sentence article dated 16 March 2026. The first sentence is the answer and the rest is about what Claude does instead.

Claude doesn't generate photos or illustrations the way image-generation tools do. Claude can, however, build diagrams, charts, and interactive visuals directly in your conversation to assist with your understanding. These are generated using HTML and SVG...

The developer documentation is blunter. The vision page, which explains how Claude reads pictures, has an FAQ entry at the bottom that most guides never open, because it sits inside a collapsed accordion.

No, Claude is an image understanding model only. It can interpret and analyze images, but it cannot generate, produce, edit, manipulate, or create images.

And the models overview says the same thing in the specification line every current model shares: text and image input, text output. There is no image output type in the Messages API, so no app built on Claude can return a picture Claude drew. When an app built on Claude shows you a picture, a second model made it.

What happened when we asked Claude for a photo?

It said no, politely and usefully. We opened a new chat in claude.ai on our Max account, with Opus 5.5 at medium effort, and typed: Make me a realistic photo of a pumpkin spice latte on a rustic wooden café table, morning light, square, for my coffee shop's Instagram feed.

We ran the same sentence through Claude Code 2.1.296 from the command line, with every MCP server switched off so nothing else could step in. Same answer, slightly more technical: I don't have an image generator built in, and I can't create a realistic photo with code. It offered to write a script for an image API if we had a key for one, and otherwise gave us a prompt to paste into ChatGPT, Gemini or Midjourney. Neither run pretended. That is worth saying, because two of the pages ranking above this one claim Claude now makes images.

What does Claude make when you ask for an Instagram post?

Something, and this is where the yes in some answers comes from. The second brief asked for a post rather than a photo: Generate an image for my coffee shop's Instagram: a square post announcing our pumpkin spice latte, $5.50, this week only. Claude treats that as a layout, and a layout is something it can write as code.

In Claude Code it wrote an SVG file, a vector drawing described in text: a cream to amber background, falling leaves, THIS WEEK ONLY, Pumpkin Spice in a serif with Latte in orange italics, a cup with a heart in the foam, a cinnamon stick, a small pumpkin and $5.50 in a brown pill. Its own summary began: I made the post, but only as an SVG. I couldn't convert it to PNG or look at it rendered. It added that Instagram will not accept an SVG and offered two ways to convert it. It also invented a tagline, Cozy up with fall in a cup, and said so.

The left image is a decent piece of clip art and it is honest about what it is. Every element is a shape Claude placed by coordinates, which is why the type is crisp and the price is right, and also why it looks like every other vector illustration: there is no light, no texture, nothing anyone would mistake for your café. Anthropic's own article on these visuals draws the same line.

Custom visuals aren't photos or illustrations. Claude builds them using HTML, the same building blocks as web pages, so they're interactive and specific to your question rather than static images.

Is Claude Design an image generator?

No, but it is the closest thing Claude has, and it is what claude.ai reached for. When we sent the post brief in chat, Claude used five tools, ran a command, and answered with a Claude Design canvas: a 1080 by 1080 artboard with a dark espresso background, a pumpkin orange circle, a line drawn cup, Pumpkin Spice Latte in heavy stacked type, a THIS WEEK ONLY tag and $5.50 at the bottom. It left [YOUR SHOP NAME] and @[yourhandle] as placeholders and told us to click them to type our own.

Anthropic launched Claude Design on 17 April 2026 as an Anthropic Labs research preview for Pro, Max, Team and Enterprise, and names social media assets in its use cases. The current help article describes it as one of the templates you can start an artifact from, usable in any chat, in Claude Code and from the Artifacts tab.

Marketing collateral: Marketers can create landing pages, social media assets, and campaign visuals, then loop in designers to polish.

What a design is made of is the important part, and the help article spells it out: a canvas of artboards, each holding layers such as text, shapes, and images. The images are ones you bring. Claude arranges them, it does not paint them, so a design with a photo of your actual latte in it is a design you uploaded a photo to.

One thing the help article does not list, and our test found: the artboard's own export menu offers PNG. The help article's export list is a zip, PDF, PPTX, Google Slides, standalone HTML and Send to Canva or Adobe. On the artboard itself, the download button opened a menu with PNG 1× at 1080 by 1080 pixels and PNG 2× at 2160 by 2160 above the PDF and HTML options. A PNG at 1080 square is a file Instagram takes as it is.

How do you get a real photo out of Claude?

By giving it a tool that calls a model which does make photos. Claude then writes the prompt, sends it, looks at what comes back and tells you what is wrong with it, which it is good at. Every route below is a variation of that, and in every one the picture is drawn by someone else's model and paid for at someone else's price.

We tried the most direct version. Same photo brief, Claude Code 2.1.296, no MCP servers, but this time with a Gemini API key in the environment and one extra sentence: use the Google Gemini API and the model gemini-3-pro-image. Claude wrote a 35 line Python script that posts the prompt to Google's generateContent endpoint with an image response and a 1:1 aspect ratio, ran it, and saved a 1024 by 1024 JPEG. The whole run took about a minute. The photo is the right half of the comparison above: a speckled ceramic cup, leaf latte art, cinnamon dusting, sun from a side window, a blurred café behind.

Then it did the part a plain image generator does not. It looked at the result and reported a problem we had not noticed: the model put the small decorative pumpkin and the cinnamon stick on top of the foam, not next to the cup as I asked. It looks cute, but it's a bit staged. It offered to run it again with them on the table. That loop, Claude writing and checking while another model draws, is the honest version of what most pages mean by Claude image generation.

If you do not use Claude Code, the same idea arrives as a connector. These are the routes the ranking pages sell, with where the picture really comes from:

Two rules from Anthropic and Hugging Face shape every connector route. Image tools are not available through the directory listing any more, and Hugging Face's own blog post carries the update, so each of these is added by URL as a custom connector. And on Anthropic's Free plan you get exactly one of those.

Update October 2025: Following an update to Anthropic's Connector Directory Policy, you will need to use the "Add custom connector" option and set the Remote MCP server URL to https://huggingface.co/mcp?login to use Image Generation and other Gradio Spaces from Claude.

Generate Images and Videos Inside Claude (Higgsfield MCP Setup), Teacher's Tech

We picked this tutorial after reading its transcript, and it shows the route honestly: Settings, Connectors, add custom connector, paste mcp.higgsfield.ai/mcp, sign in with a free Higgsfield account. A bakery image comes back, and the presenter says it was created with Nano Banana by default before switching to GPT Image 2. Two things to know: the video is sponsored by Higgsfield, which it says at the 30 second mark, and nothing in it is drawn by Claude.

What do the other pages about Claude image generation get wrong?

Mostly the headline. The pages that rank for this question either quote Anthropic and agree, or sell a connector and call its output Claude's. Here is what each says against what Anthropic's pages say.

Will Instagram label an image made this way as AI?

Probably, if it is the photo. We counted 24 occurrences of the C2PA marker in the raw JPEG that Gemini returned to Claude Code, which is the signed manifest that says a file was made by AI. Meta reads that standard to apply the AI info label, which we covered with Meta's own words in why Instagram says AI info on your post, and TikTok does the same, see the TikTok AI-generated label. The SVG is a different case: it is a drawing written as code, with no manifest in it for a platform to read. We could not download the Claude Design PNG in our test browser, so we have not checked that one.

Whether that label matters is your call. Meta has not published what it does to reach, and for a picture that is not of your real product it is arguably the honest outcome. For a photo meant to show your real product, a phone picture of the real thing is better on every count.

So which should a small business use?

Use Claude for what it is good at and let it pick the tool for the rest. If the post is type, a price and a photo you already have, Claude Design will lay it out and export a 1080 pixel PNG, and that costs nothing beyond your plan. If you need a photo you do not have, Claude will write a better prompt than most people do, and you can paste it into Gemini or ChatGPT yourself, or connect one of the image tools and let Claude paste it for you. If you already use Claude Code, a key and one sentence got us a usable photo for about 13 cents.

What none of these routes do is the whole job: the picture, the text set inside it, your brand, and getting it onto the feed. For the last step, see how to connect Claude to Instagram, and for the writing, our tested Claude prompts for Instagram posts. If you want the image model's side of this, the prompts we ran in Gemini prompts for poster design and for product photography are the ones that held up.