Skip links

Can Claude generate images? Charts and icons yes, photos no

Claude images typed at a terminal prompt on a cream AIWIZ card, with coding task rows below
Claude has no image generator. It draws with code, and the Claude apps hand you the result as a PNG, SVG, PDF or HTML file.

Claude has no image generator, so it cannot produce a photo or a painted illustration the way ChatGPT, Gemini or Grok can. It can draw with code: SVG illustrations, charts, diagrams, icons and social graphics, which the Claude apps turn into PNG, SVG, PDF or HTML files you can download.

We tested this on 13 November 2026 in Claude on the web, in one account on the Ultra plan, with Haiku 5.5, Opus 5.5 and Fable 5.2. Asked for a photorealistic picture of a red bicycle against a dry-stone wall, all three said they could not make one, and two of them wrote a prompt to paste into an image generator instead. Asked to draw the scene, all three returned an SVG and a PNG file, in times from about 70 seconds to 34 minutes.

For a UK marketing team, that makes Claude a good tool for charts, diagrams, icons, simple social graphics and image prompts, and the wrong one for photography, product shots or lifelike people.

The answer in brief

  • Claude’s models take text and images in and return text. Anthropic’s help centre says Claude “doesn’t generate photos or illustrations the way image-generation tools do”.
  • Claude draws by writing code. In the Claude apps it runs that code and hands you the file, on every plan including Free.
  • In our tests on 13 November 2026 it made an illustration, a chart, a flowchart, an icon and a social graphic, and it cropped, resized, converted and recoloured a photo we uploaded. It declined to add a croissant that was not in the photo.
  • Anthropic says it does not currently focus on image or video generation models. Animations are drawn with code, and voice mode speaks Claude’s replies in preset voices.
  • Anthropic’s connector directory generally keeps out AI image, video and audio generators. Design tools such as Canva, Adobe and Figma are listed, and a custom connector can attach an image service on your own account.
  • ChatGPT, Gemini, Grok, Microsoft Copilot and Meta AI each have image generation built in.
  • Our advice: use Claude for charts, diagrams, icons, layouts and image prompts, and check each result against your brief. Use an image model or a photographer for photos.

Can Claude generate images?

No, if you mean a photo or a painted illustration made from a prompt. Claude has no image generation model, and nothing built into the Claude apps paints a picture. It can still hand you an image file, because it draws with code.

Anthropic’s documentation says so at three levels. The page for each current model, Claude Opus 5.5 among them, lists its input and output as text and images in, text out. The vision guide calls Claude “an image understanding model only”. The help centre puts it in plainer words: Claude “doesn’t generate photos or illustrations the way image-generation tools do”.

On 13 November 2026 we asked for exactly that in Claude on the web, on an Ultra plan: “Generate a photorealistic image of a red bicycle leaning against a dry-stone wall in the Yorkshire Dales at golden hour.” We sent the same prompt to three models, each in a new chat.

  • Opus 5.5 at medium effort checked its tools, then answered: “I can’t make a photorealistic image here. Nothing in this session generates photographs.” It offered to draw an SVG illustration and wrote a prompt for an image generator, naming Midjourney, DALL-E, Imagen and Firefly.
  • Haiku 5.5 at medium effort said “I can’t generate images here”, listed what its tools could do (search for existing photos, draw SVG or HTML graphics, build designs) and wrote a prompt of its own.
  • Fable 5.2 at max effort searched Anthropic’s connector directory for an image tool and found none, adding that its search covered only part of the directory. It then offered a prompt, an illustration or a search for real photographs.

Each answered within about half a minute, and none produced a file.

Claude's reply to a request for a photorealistic image: it says it cannot make one and writes a prompt for an image generator
Figure 1. Asked for a photo, Claude says it cannot make one and writes the prompt for a tool that can.

That prompt is worth keeping. It names the lens, the light and the setting, and Opus 5.5 added a negative prompt because, in its words, image generators “often get bicycle frames and wheels wrong”.

With web search on, Claude can also find existing images on the web and show them in the chat, each with a link to its source. Those are other people’s pictures, with their own usage rights.

Why does Claude not generate images?

Anthropic says it has chosen to put its computing power elsewhere. In the Form S-1 it filed with the US Securities and Exchange Commission in October 2026, the company says it does not currently focus on developing image and video generation models, despite their commercial appeal, and that the compute goes to coding, agentic skills and safety research instead.

Its chief executive had given another reason earlier. On stage at the Cerebral Valley AI Summit in November 2024, Dario Amodei told Newcomer that image and video models feel more like a direct replacement for human creative work and raise “a wider range of safety issues”. He called that one reason, and not the only one, why Anthropic had been less involved in them.

What can Claude make instead of AI-generated images?

Claude makes pictures that can be described in code: illustrations, icons, charts, diagrams, page layouts and simple animations. SVG is a file format that describes shapes and text for software to draw, and a chart can come from a short plotting script. Claude writes that code, and in the Claude apps it also runs it and returns the finished file.

Code can keep the text and shapes in a picture editable. In our tests the chart’s labels matched the figures we supplied and the social graphic had no spelling mistakes. Check the wording and the numbers anyway, and look at the whole picture again after any change.

What Claude can make, how it makes it and the file you get

If the table extends beyond the screen, scroll sideways to view all columns.

What you want How Claude makes it Where it works File you get
A diagram, flowchart or chart inside the chat How Claude makes itBuilds a custom visual from HTML and SVG Where it worksWeb and desktop apps on every plan (beta) File you getCopy as image, SVG or PNG offered in our test
An illustration, icon or logo draft How Claude makes itWrites SVG code and converts it to PNG if asked Where it worksApps with code execution and file creation on, every plan File you getSVG, PNG
A chart or data graphic from your figures How Claude makes itWrites and runs a plotting script Where it worksWeb, desktop and mobile apps, every plan File you getPNG, and PDF, Excel or PowerPoint files
A social graphic, poster or mock-up How Claude makes itLays it out on an editable canvas in Claude Design Where it worksEvery plan, with code execution and file creation on File you getPNG, PDF or HTML offered in our test
A short animation How Claude makes itDraws the frames with code and encodes them, or uses Claude Motion Where it worksApps with file creation on; Motion is a beta for Team and Enterprise File you getGIF, MP4
A changed version of your own photo How Claude makes itRuns image-editing code on the file Where it worksApps with code execution and file creation on File you getWebP in our test

Sources: Anthropic’s help centre on custom visuals, creating and editing files, artifacts, Claude Design and Claude Motion, and the file types returned in AIWIZ’s own tests, described below.

We ran each row except Motion on 13 November 2026, in Claude on the web with Opus 5.5 at medium effort unless we say otherwise. The photo request and the drawing request ran once on each of the three models, and every other prompt ran once. We did not test the Free plan, the mobile apps, a connected image service or the API, and we looked at the files in the chat without downloading them.

An illustration. After the photo request we asked each model: “Draw the scene as an SVG illustration instead, 1200 by 630 pixels. Then give me the same picture as a PNG file.” Haiku 5.5 returned both files in about 70 seconds and remarked that its wall “reads more like brickwork than dry stone”. Opus 5.5 took about two and a half minutes. It rendered the picture, looked at its own PNG, saw that the wall’s shadow fell on the wrong side for a sun on the left, and corrected it before sharing the files. Fable 5.2 at max effort worked for about 34 minutes, rendering its drawing again and again, reading crops of it and refining details.

Fable 5.2’s drawing has the most detail of the three. For this brief we would pick the Opus 5.5 result: a clean illustration in under three minutes. One run each, at different effort settings, cannot show which model is best for design work.

The same bicycle and dry-stone wall scene drawn in code by Claude Haiku 5.5, Opus 5.5 and Fable 5.2, in rising detail
Figure 2. One request drawn by three Claude models: the most detailed drawing took by far the longest.

A chart. We gave Claude monthly search volumes from Ubersuggest and asked for a grouped bar chart as a 1200 by 630 PNG, with each bar labelled and the source underneath. The file arrived in under 80 seconds and matched the figures we supplied, among them 1,000 UK searches a month and 8,100 in the US for “can claude generate images”. A portrait version for phones took about 50 seconds more.

Grouped bar chart Claude made from figures we supplied: UK and US monthly searches for three phrases, every bar labelled
Figure 3. A chart Claude plotted from figures we supplied and returned as a PNG file.

An icon. Asked for “a simple icon for a bakery: a wheat sheaf inside a circle, in two colours, as an SVG with a transparent background”, Claude returned a 512-pixel SVG in brown and gold. It rendered a preview to check its own drawing and warned that the thin inner ring would disappear at 16 pixels.

A social graphic. In Claude Design we asked for a square graphic, 1080 by 1080 pixels, announcing that a village bakery now opens on Sundays, with warm colours, large type, no photographs and no brand name. It was on the canvas in about 90 seconds with every word spelled correctly. The canvas has tools for changing the text and shapes by hand, which we did not try.

A wheat sheaf icon Claude wrote as SVG, and a square bakery opening-hours graphic it laid out in Claude Design
Figure 4. An icon written as SVG and a social graphic laid out in Claude Design, each from a two-sentence request.

How do you get an image file out of Claude?

Ask for the format and size you need, for example a PNG at 1200 by 630 pixels. How you save the picture depends on where Claude put it.

  • A file in the chat. With code execution and file creation on, Claude attaches the file to its reply with a download button. Anthropic says the feature is on by default on Free, Pro, Max and Ultra plans, and that owners control it on Team and Enterprise plans.
  • A visual drawn inside the reply. Our flowchart appeared in the conversation itself. A menu at its top right offered “Copy as image”, “Download as SVG” and “Download as PNG”. Anthropic’s help page lists slightly different options (an SVG or HTML download, and saving the visual as an artifact) and warns that these visuals are temporary unless you keep them. They do not render in the iPhone and Android apps.
  • A design on the canvas. In Claude Design, the Export tab offered HTML, PDF and PNG downloads for our one-page graphic. Anthropic’s help page also lists a .zip file, PowerPoint and Google Slides, and designs can be sent on to tools such as Canva, Adobe, Miro and Wix once you connect them.
The save menu on a flowchart in a Claude chat (copy as image, SVG or PNG) and Claude Design's Export tab (HTML, PDF, PNG)
Figure 5. The two save menus we opened: the options on a flowchart drawn in the chat, and the Export tab in Claude Design.

If the image is going on a website, ask for PNG or WebP. WordPress, for one, does not accept SVG uploads by default. If you need it larger, ask: after the drawing test, Opus 5.5 offered a sharper PNG at twice the size.

Can Claude edit images?

Yes, with code. In our test Claude cropped, resized, converted and recoloured a photo, and declined to add a realistic object that was not in it.

We uploaded a 600 by 400 pixel photograph of a dark red-brown coffee cup on a matching saucer (a public-domain picture by Rachel Michetti from the scikit-image sample library) and asked Claude to crop it to a square around the cup, resize it to 1080 by 1080 pixels and save it as a WebP file. In the same message we asked it to “change the picture itself: make the cup red”. It did all of that in about 90 seconds, turning the cup a bright red and leaving the saucer alone. For the recolour it marked out the cup’s outline itself, because the cup and saucer were the same colour, and it told us that enlarging so small a photo about three times had left the result soft.

Then we asked it to add a croissant on the table “so that it looks like part of the original photo”. Claude declined: “I can’t add a croissant that would pass as part of the original photo.” It said it had code that edits pixels and no image generator, and offered to cut a croissant out of a photo we supplied and blend it in.

A coffee cup photo, Claude's square crop of it and the same crop with the cup recoloured a brighter red, all done with code
Figure 6. Claude cropped the photo to a square and recoloured the cup with code.

Cropping, resizing, converting formats and recolouring are jobs for code, and our test covered each of them on one photograph. Code can also add text or a logo and combine pictures you supply, which we did not test. Adding a realistic new object or a new background scene is a job for a photo editor or an image generation tool. Anthropic’s vision guide says Claude cannot “generate, produce, edit, manipulate, or create images”, which describes the model. The edits in our test were made by code that Claude wrote and ran.

Can Claude read images?

Yes. Every current Claude model takes images as input, and the apps accept JPEG, PNG, GIF and WebP files of up to 8,000 by 8,000 pixels, 20 files to a chat. Claude also reads the charts and pictures inside a PDF of up to 100 pages. Anthropic suggests images of at least 1,000 by 1,000 pixels for the best results.

That makes Claude useful for the reverse job: describing a photo, pulling the numbers out of a chart, checking a design against a brief or reviewing a screenshot of an advert. It used the same ability on its own work in our tests, when it looked at its drawing and moved the shadow. One caution from Anthropic’s vision guide: Claude cannot determine whether an image is AI-generated, so do not use it as a detector.

Can Claude create videos, animations or GIFs?

Claude can make simple animations with code, and it cannot generate video footage. It has no video model.

We asked for “a five-second animation of a bar chart whose four bars grow from zero to 3, 5, 8 and 13”, as a GIF and as an MP4 file. Claude returned both and reported each as exactly five seconds long, after rebuilding the GIF because its first attempt ran for 4.6 seconds. We did not download or play either file. Asked how it had made the MP4, it said it used FFmpeg, which was already installed in its workspace. Anthropic’s list of the libraries in the API’s code execution sandbox does not include FFmpeg, so that result holds for the Claude app and may not hold through the API.

Anthropic also offers this as a product. Claude Motion, a beta for Team and Enterprise plans, builds an animation as code and exports it as MP4. Anthropic’s own description sets the limit: Motion “doesn’t use a video generation model, so it doesn’t create realistic footage or people”. Our Ultra account did not offer Motion. Claude Data, on Pro plans and above, can turn figures from a connected data source into narrated short videos. We did not test it.

Can Claude generate audio, speech or music?

Only in a limited way. Claude’s models return text, voice mode speaks its replies aloud, and in our test its workspace made two audio files with code.

Voice mode is a beta on every plan, in the mobile and desktop apps and on the web. You speak, and Claude answers aloud in one of a small set of preset voices. Anthropic’s help page notes that the Fable models are not available in voice mode.

For files, we asked Claude to read a sentence aloud and to compose a five-second tune. It returned two MP3 files in about a minute and said it had made the speech with Piper, an open-source text-to-speech engine that it installed during the chat. We did not download the files, so we have not heard either one. We would treat that as a curiosity: it depends on the workspace being allowed to install software, and the voice belongs to a third-party engine with its own licence. For a voiceover that will be published, use a dedicated speech tool.

Can connectors or MCP servers add image generation to Claude?

Yes, through a connector to another company’s service. Anthropic’s directory generally does not list AI image generators, design tools are listed, and a custom connector can attach an image service.

Anthropic’s software directory policy says that, unless Anthropic permits it in writing, the directory does not accept software that “uses AI models to generate images, video, or audio content”. It names text-to-image, image editing and people generation, along with video models, text-to-speech and music models. The exception is “design-focused image software” that makes visual aids such as slides, diagrams, charts, mock-ups and logos, as long as the developer does not offer standalone image generation inside Anthropic’s products. Anthropic adds that it may revisit these restrictions.

Canva’s connector can generate and export designs in your Canva account. Adobe’s edits photos one at a time or in bulk, removes backgrounds and resizes images and video through Adobe’s own tools. Figma and Gamma are listed, as are the stock libraries Unsplash and Shutterstock. We looked in the directory for more than a dozen image, video and music generators by name, Midjourney, Ideogram, Stability, Kling and Suno among them, and none was listed.

Outside the directory, custom connectors are available on every plan, with Free limited to one. On Team and Enterprise plans an owner has to add the connector before anyone can use it. They connect Claude to a remote MCP server that is reachable over the public internet, and that server can belong to an image service. MCP is the open standard that lets an AI application connect to another service’s tools and data. Hugging Face, for example, tells its users to add its server as a custom connector to use image generation from Claude, following a change to Anthropic’s directory policy. Anthropic’s warning is plain: custom connectors link Claude to services “that haven’t been verified by Anthropic”. We did not connect or test an image service for this guide.

Whichever route you take, the picture is made by the other company’s model, so its quality, its price and its licence terms are theirs. You usually need an account with that service, and sometimes credits. Claude’s part is the brief: it writes the prompt, calls the tool and can check the result.

Claude Code and Claude Cowork follow the same rule. Claude Code can also run commands, so it can call an image service you already have keys for, and the image still comes from that service. Developers using the API get text back from the model. Its code execution tool can save a chart or drawing as a file in a sandbox with no internet access, and the Files API downloads that file.

One more caution. Not every site with Claude in its name is Anthropic’s. One we read sells image generation, says it is not affiliated with Anthropic and names other companies’ image models. Anthropic’s own apps are Claude on the web at claude.ai and the Claude desktop and mobile apps. Before you pay or upload client material anywhere else, check who runs the service and which model makes the pictures.

Is Claude or ChatGPT better for image generation?

For photos and painted illustrations, ChatGPT, because it has image generation built in and Claude does not. The same goes for Gemini, Grok, Microsoft Copilot and Meta AI. We did not compare the picture quality of any of them.

Which AI assistants have image generation built in

If the table extends beyond the screen, scroll sideways to view all columns.

Assistant Image generation built in? What its maker says
Claude (Anthropic) Image generation built in?No What its maker saysBuilds diagrams, charts and visuals from HTML and SVG, and reads images you upload
ChatGPT (OpenAI) Image generation built in?Yes What its maker saysImage creation is on every plan, from “Limited” on Free to “Unlimited and faster” on Pro, subject to abuse guardrails
Gemini (Google) Image generation built in?Yes What its maker saysThe Gemini app creates and edits images with its Nano Banana models, free of charge, with its most advanced image model on a Google AI plan
Grok (SpaceXSI) Image generation built in?Yes What its maker saysGrok Imagine covers image generation, image editing and video generation
Microsoft Copilot Image generation built in?Yes What its maker saysCopilot has “a built-in ability to create images from text-based prompts” and can change an image you upload
Meta AI Image generation built in?Yes What its maker saysMuse Image makes images in Meta AI, “free for everyday creation”

Sources: Anthropic’s help centre, OpenAI’s ChatGPT plans page, Google’s Gemini image page, SpaceXSI’s developer documentation, Microsoft Support and Meta’s Muse Image announcement.

OpenAI, Google, SpaceXSI and Meta each offer a separate image model: GPT Image, Nano Banana, Grok Imagine and Muse Image. Gemini 4 Carbon and Gemini 4 Hydrogen write text only, and the pictures in the Gemini app come from Nano Banana. Meta AI’s pictures come from Muse Image, which our guide to Meta’s Muse models covers. Anthropic has not released an image model, so Claude needs an outside image tool to make a new photorealistic picture.

The vendor pages do not all say which countries each feature covers. Check UK availability and the licence terms before you plan a campaign around one.

Which tool should a UK marketing team use for which job?

Use Claude where the picture is built from shapes, text and data, and an image model or a camera where it has to look real.

  • Charts and data graphics: Claude. It plotted the figures we gave it with matching labels. Compare any chart with its source data before you use it.
  • Diagrams, flowcharts and process maps: Claude, drawn in the chat or on a whiteboard.
  • Icons and logo drafts: Claude for a first SVG, then a designer for anything that becomes the brand.
  • Social graphics, posters and one-pagers made of text and shapes: Claude Design, or Canva through its connector if your templates live there.
  • Photos, product shots and lifestyle images: an image model, a stock library or a photographer. Claude can write the prompt.
  • Photo edits: Claude for cropping, resizing, format changes and simple recolouring. A photo editor or an image model for retouching and for adding or removing things.
  • Video and voiceover: Claude for animated charts and explainers drawn in code. A video or speech tool for footage and narration.

The two kinds of tool work well in sequence. Ask Claude for the image prompt, with the format, the light, the composition and a list of what to leave out. Run it in the image tool. Then bring the result back to Claude to check it against the brief, describe it for alt text or resize it for each channel. Our guide to prompting Claude covers how to brief it.

What are the limits of images made by Claude?

The pictures Claude draws look drawn, and that is the limit that matters most. The others are practical.

  • Realism. What Claude drew in our tests was illustration and layout. For a new lifelike picture of a person or a product, use an image generation tool or a camera. Claude can edit a photo you supply.
  • Time and usage. Our drawings took from about 70 seconds to 34 minutes. Anthropic says file creation uses more of your plan’s limit than ordinary chat, and the model selector labelled Fable 5.2 at max effort “3.5× or more usage”.
  • Protected work. Anthropic’s published system prompt for Claude Fable 5.2 tells the model not to reproduce a specific artwork, cover, poster, logo or product design, and not to draw a known character or mascot at all. It allows original designs, generic subjects, public-domain works and your own artwork or logo. The rule covers anything it draws with code, SVG included.
  • Labelling. Anthropic says that when Claude generates a supported file type, such as a PNG or JPEG, it attaches Content Credentials, signed metadata showing that the file was processed by Claude. Support depends on the model, the platform and the feature, and converting, re-saving or screenshotting a file can strip the metadata. We did not download a file to check.
  • Phones. Visuals drawn inside a reply do not render in the iPhone and Android apps. File creation and voice mode do work there.

Need help choosing AI tools for your marketing images?

AIWIZ can test the options against your own briefs and brand assets, and set up the workflow that gets from a prompt to a finished file. Talk to us about your AI or automation project.

Frequently asked questions

Will Claude get image generation?

Anthropic has not announced an image generation model for Claude. In November 2024 Dario Amodei said the company had "no current plans" for image or video models and that he never wanted to say never, and its Form S-1 of October 2026 says it does not currently focus on them. Its software directory policy says the limits on outside image, video and audio generators may be revisited, which is a rule for other companies' tools.

Is Claude Design an image generator?

No. Claude Design arranges text, shapes and the images you give it on an editable canvas. Anthropic says it "doesn't have an image model" and is not built for image generation. In our test it built a social graphic from type and shapes alone.

Explore
Drag