Models · Tested
Can Claude Generate Images? We Tested Opus 5.5 Two Ways
Short answer: no, not the way an image model does. Every current Claude model, from Fable 5.1 and Opus 5.5 to Sonnet 5.5 and Haiku 5.5, takes text and images in and gives text out. Claude can still get you a picture in two ways: it can write the picture as code, or it can direct an image model that paints it.
We tested both routes with Claude Opus 5.5 on one fictional job for Rook & Rabbit, a board-game café. On September 27, 2026, in claude.ai, Opus drew a 32 × 32 mascot sprite as SVG. On October 15, 2026, in Ciyo, whose agent runs on Opus 5.5, it made the café's Friday game-night post with GPT Image 2.5.
What Claude can and cannot make
Anthropic's model overview lists text and image input and text output for all current Claude models. Its Help Center is plainer: Claude does not make photos or illustrations the way image tools do. It can build diagrams, charts and other visuals from HTML and SVG, and it can read and describe the images you upload.
So when you ask Claude for a picture, what comes back depends on the picture. Flat shapes come back as code that a browser draws. A photo or a painted scene does not come back at all; you get a description, or a prompt you can give to an image model.
| You ask for | Claude gives you | Use it for |
|---|---|---|
| A logo, icon or simple graphic | SVG code that a browser draws | Flat shapes you can scale and edit |
| Pixel art | SVG or canvas code that places each pixel | Sprites that need an exact grid |
| A chart or diagram | HTML or SVG built in the chat | Data and process visuals |
| A photo or painted scene | A description or prompt, not an image | Handing the brief to an image model |
| Changes to your photo | Feedback on the photo, not a new version | Deciding what to fix in an image tool |
Why your feed says Claude can now make images
On October 11, 2026, a short video titled “Claude can finally generate images” showed the Claude app answering “Generate an image of a red car.” By October 15 it had passed 1.7 million views. A Community Note under it says Claude has no native image generation and that the car is SVG code for a basic vector graphic.
The poster replied that SVG is an image format, and that is true. It is also the whole story: the car is a drawing that Claude wrote as code. That route is good for some jobs and wrong for others, which is what the two tests below show.
Posted 2026-10-11 (UTC). The video shows the Claude app drawing a red car; readers added a Community Note explaining that it is SVG code, not native image generation. The post is not about Ciyo, and the tests below are our own.
Test 1: Opus 5.5 draws a pixel-art mascot as code
Pixel art is the best case for the code route, so we started there. We asked Opus 5.5 at the Medium effort level in claude.ai for one sprite with hard limits we could check afterwards: a 32 × 32 grid, one square per pixel, at most eight colours, a transparent background and no text.
Draw a 32 x 32 pixel-art sprite of the mascot for Rook & Rabbit, a board-game cafe: a white rabbit hugging a black chess rook, facing right, on a transparent background. Draw it as one SVG with viewBox="0 0 32 32" and shape-rendering="crispEdges", where every pixel is one 1x1 <rect> on the integer grid. Use at most 8 colours, including a dark outline colour. No text. Reply with the SVG code only, in one code block.It looked at its own drawing four times
Before it answered, Opus 5.5 turned its SVG into a picture and looked at the result. The step list shows four rounds: a first draft, a revised sprite, a patch for gaps and paws, and a hand-composed version, each followed by reading the rendered file. The whole answer took a little over a minute.
That loop is the main reason the result holds together. A model that writes code for a picture without seeing it tends to leave gaps and stray pixels.
What we measured
We copied the SVG, 26,444 bytes, and counted it with a script. It holds 629 squares, every one exactly 1 × 1 on the integer grid, with no duplicates and nothing outside the 32 × 32 frame. It uses 6 colours, under the limit of 8, and the unused pixels are transparent.
Compare that with “pixel art” from an image model. In our photo-to-pixel-art test the same day, a generated image that looked like pixel art held 173,508 distinct colours and blocks of uneven size. Code gives you the grid by construction.
| Requirement | Asked | Delivered |
|---|---|---|
| Grid | 32 × 32 | 32 × 32, 629 filled pixels |
| Pixel shape | 1 × 1 squares on whole numbers | All 629, 0 off the grid |
| Colours | At most 8 | 6 |
| Background | Transparent | Transparent |
| Text | None | None |
Where the code route falls short
Look at the arms. At this size the rabbit's hug reads as two white bars across the rook, and most of the rook hides behind the body. A pixel artist would cheat the pose to keep both characters readable; the model kept to the literal brief.
The file is also heavy for what it is: 26,444 bytes of SVG for a picture that saves as a PNG of a few hundred bytes. Export it to PNG at 1× and at a whole-number scale before you use it in a game or on a website. And no amount of code gives you light, fur or a lived-in café; for that you need an image model.
Test 2: Opus 5.5 directs an image model in Ciyo
Ciyo's agent runs on Claude Opus 5.5. It does not paint the picture itself: it reads the brief, writes the image prompt and calls an image model, then puts the result on the canvas. On October 15, 2026, we sent it one message on dev.ciyo.ai, our test site, from a fresh project.
The agent replied that the brief was clear enough to go straight to one square post, and called GPT Image 2.5 Flare. A 1024 × 1024 image appeared on the canvas; the whole turn, including the check below, took about 40 seconds. Our run record for that turn names the model as anthropic/claude-opus-5.5.
Make a square Instagram post for Rook & Rabbit, a small board-game café. It announces Friday Game Night at 7 pm. Show a white rabbit hugging a black chess rook on a café table with board games and warm evening light. Put the words "Friday Game Night · 7 pm" on the image.Then it checked the picture
Claude reads images as well as text, and the agent used that. After the image arrived it said it would check that the text rendered correctly, opened the picture, and reported back: the headline reads “Friday Game Night · 7 pm” across the top, and a small “Rook & Rabbit” line sits at the bottom. It offered a 2K version for 1080 × 1080 posting, or a variation with the café's real logo.
Compare the two rabbits. In the poster the hug reads at a glance, with fur, a glossy rook, dice, cards and a latte in string-light glow. The sprite could not show any of that. Two small things to check before you post: the image model added the “Rook & Rabbit” line without being asked, and the agent's check is a reading, not a guarantee, so look at the lettering yourself. The canvas showed 6,066 credits as the turn started and 6,059 when it ended, on our test account.
Which route to use
Pick the route by the picture, not by the model. Claude is useful on both routes: as the author of exact code on one, and as the planner and checker on the other.
There is a third way you may meet: a connector, or MCP server, that lets Claude call an outside image model from the chat. The picture then comes from that model, so its quality, licence and price are that model's, not Claude's. Claude Design, which Anthropic launched on April 17, 2026, lays out prototypes, slides and one-pagers and exports them as PDF, PPTX, HTML or to Canva; Anthropic does not present it as a photo generator.
| You need | Route | Why |
|---|---|---|
| An exact sprite, icon or simple logo | Ask Claude for SVG | Every shape lands where the code says, and you can edit it |
| A chart, diagram, slide or one-pager | Claude chat or Claude Design | Built from HTML and SVG, so text stays sharp and editable |
| A photo, poster, product shot or painted scene | An image model | Light, texture and lettering that code cannot fake |
| A finished post from a plain-language brief | An agent that plans with Claude and calls an image model, such as Ciyo | One message: it writes the prompt, makes the image and checks it |
Claude and images
Can Claude generate images?
Not natively. All current Claude models give text out. Claude can write code, such as SVG, that draws a picture, and it can direct an image model through a tool or an agent that makes the picture for it.
Can Claude Opus 5.5 generate images?
Not as an image model does. In our test on September 27, 2026 it drew a 32 × 32 sprite as SVG and checked its own renders. In Ciyo on October 15, 2026 it made a photographic poster by calling GPT Image 2.5.
Why did a viral video show Claude making a picture of a red car?
Claude wrote the car as SVG code and the app drew it. A Community Note on the October 11, 2026 post says the same: this is a code-drawn vector graphic, not native image generation.
Is Claude's pixel art exact?
In our one test, yes: 629 one-pixel squares on a 32 × 32 grid, 6 colours, no stray or oversized squares. The pose was weaker than a pixel artist's.
Can I use Claude Opus 5.5 in Ciyo?
Yes. As of October 15, 2026, Ciyo Agent runs on Claude Opus 5.5. It plans the work and calls image models such as GPT Image 2.5 to make the pictures on your canvas.
What does Opus 5.5 cost?
Anthropic lists $4 per million input tokens and $20 per million output tokens on its API. In claude.ai it is part of the paid plans.
Ask Ciyo's agent for the picture
Describe the post, poster or product shot. Ciyo Agent plans it with Claude Opus 5.5, makes it with image models such as GPT Image 2.5 and checks the result.