Ideogram Vector 1.0: AI Images You Can Edit as SVG

Ideogram announced Ideogram Vector 1.0 on October 22, 2026, calling it the company's first vector model and listing it as live in the Ideogram app and the API the same day. It generates an SVG file from a text prompt, optionally guided by reference images, and it can convert an existing raster image into an SVG. Prices as read on October 26 run from $0.13 to $0.32 per generated file and $0.10 per vectorised image.

Most AI image models produce a grid of pixels. That is fine for a social post and wrong for a logo, a label, a sign or anything that has to be printed at a size nobody chose yet. Designers have worked around it by tracing AI output into vectors by hand or with an auto-tracer, and both routes cost time and lose fidelity. A model that writes the SVG directly removes that step, which is why the deliverable, not the image quality, is the story here.

This post is for the design and web teams deciding whether to route logo, illustration and print work through it. It covers what shipped, what an editable SVG does and does not give you, the per-file pricing and how the API behaves, then a set of checks to run on any file before it goes to a client or a printer. We have not run the model ourselves; every capability claim here is Ideogram's, and we say so where it matters. For how we specify editable deliverables generally, our 3D asset delivery reference applies the same thinking to meshes and scenes.

The announcement thread makes four claims in sequence: it is Ideogram's first vector model; it generates editable SVGs with smooth gradients and sharp typography from a prompt, or vectorises any image; each design is one clean SVG editable in Illustrator, Figma or code; and it is live in the app, the API and on partner platforms, which the thread does not name. The model page adds the identifier, ideogram-vector-1, two API paths and the price table.

Two things it is not. It is not an open-weight model: Ideogram's open release remains Ideogram 4.0, and nothing on the vector pages mentions weights. And it is not a tracing filter bolted onto a raster model; the text-to-vector endpoint takes a prompt and returns an SVG, with no intermediate image you can see. The image-to-vector endpoint is the tracing path, and it is priced and documented separately.

A raster image is a fixed grid of coloured pixels. A vector file is a list of shapes: paths with coordinates, fills, strokes and gradients that a renderer draws at whatever size it is asked for. The practical consequences are the three the model page leads with: the artwork stays sharp from a label to a poster, colours can be changed without regenerating, and individual shapes can be selected and refined in a vector editor. Those are properties of the file format, and Ideogram's contribution is producing the format directly.

Here is the workflow that breaks without it. A marketing team generates a logo concept with a raster model, the client approves it, and the sign company asks for a vector. Someone auto-traces the PNG, the gradients turn into banded stripes, the lettering edges go lumpy, and a designer spends an afternoon rebuilding the mark by hand. The concept cost pennies; the conversion cost a day. A native SVG is meant to delete that day, and whether it does is the only question that matters for a studio.

The second consequence is quieter. When the client comes back with "can the blue be our blue", a raster workflow regenerates and hopes the composition survives. A vector workflow edits one fill value. Ideogram's thread puts this as swapping the palette and having the gradients follow. If it holds, revision cost drops from a new generation and a new review to a two-minute edit, and that is where the economics of the deliverable change.

An SVG is editable in the sense that any text file is editable. Whether it is usefully editable depends on how the shapes are built, and Ideogram publishes nothing about that: no path counts, no file sizes, no statement about how a gradient is encoded or whether overlapping shapes are merged. Three gaps in particular should be assumed until you have opened a dozen files.

First, text. Ideogram's thread describes its typography as drawn as clean shapes for print and digital. That is a feature claim and also a limit: lettering arrives as outlines, not as a text element you can retype. The failure case is the obvious one. A client approves a wordmark, then renames the product. In a designer-built file you change the string; in an outlined file you regenerate or redraw. The model page's gallery leads with typography, logos and labels, so for the work this model targets it is the common case, not the edge.

Second, topology. A vector editor can select any shape, but a design that arrives as hundreds of tiny overlapping fragments is editable only in theory. Auto-tracers are notorious for this; whether a generative model does better is something Ideogram has not published and we found no public measurement of. Third, print. SVG colours are RGB. A printer works in CMYK or spot colours, and a brand blue defined in one space will not match in the other without conversion. None of this is a criticism of the model. It is the list of things the word "editable" leaves for you to do.

The price table has one shape worth noticing: references are free at low and medium quality and cost 60% more at high. Low and medium cost $0.13 and $0.16 either way; high is $0.20 prompt-only and $0.32 with references. The generate endpoint defaults quality to medium, so an integration that never sets the field pays $0.16 a file and never produces the top tier.

Put those numbers against the work they replace and the scale of the change is clear. A hypothetical logo exploration of 40 concepts at medium quality costs $6.40. Pushing the five shortlisted concepts to high quality with a brand-reference image costs another $1.60. Vectorising the client's existing raster mark so it sits alongside them costs $0.10. The whole exploration is under $10 in API spend; the expensive part is the designer's hour reviewing it, which is the correct place for the cost to be. The arithmetic is ours on Ideogram's list prices.

The failure case is paying for quality you cannot use. The high tier is the one you want for a final mark; it is a waste on exploration, where the question is composition, not edge quality. Set quality per stage. And use the dry-run flag the API offers: it validates and prices the exact request without generating or billing, which is the cheapest way to catch a batch job that accidentally sends eight high-quality images with references on every call.

Both vector endpoints are asynchronous. A request returns a generation ID, and the result is collected by polling a shared generations endpoint or by receiving a webhook. The webhook delivery is signed with Ed25519 against public keys Ideogram publishes, must be HTTPS, and rejects private, loopback and cloud metadata hosts. Outputs are private by default, and enterprise outputs are always private. All of this is from the developer reference as read on October 26.

The generate endpoint takes a multipart form: a prompt of up to 9,500 characters, optional reference images either as uploads or as identifiers of assets already in your account, an aspect ratio that defaults to auto, a quality that defaults to medium, one to eight images, and a seed. The reference documents the first image as the primary source and any others as additional references, which matters when you pass a brand mark plus style samples: put the mark first.

The rejection codes are worth reading before you write the retry logic. A 402 or 429 carries a reason: insufficient funds, a subscription requirement, a daily limit, a priority-credit requirement, a feature limit, or an in-flight limit that comes with the maximum number of concurrent generations your account is allowed on the queue it was routed to, fast or slow. The failure case is a batch script that treats every non-200 as transient and hammers a daily limit until the key is throttled. Branch on the reason field.

The second endpoint takes a JPEG, PNG or WEBP of up to 50 MB and returns a generation ID for an SVG, with no prompt. Its one setting is a mode with three values that the reference describes plainly: simple uses fewer shapes, detailed preserves fine detail, and auto chooses between them for the image. At $0.10 a file it is the cheapest operation Ideogram offers, and for an agency it may be the most used: every client has a logo that exists only as a JPEG from 2014.

The model page's examples are flat artwork: lettering over cat eyes, a pop-art portrait. That is the right scope. A photograph vectorised in detailed mode becomes thousands of shapes that scale cleanly and edit terribly. That is the usual result of tracing a photograph, and nothing Ideogram publishes suggests otherwise here. The use case is artwork that was always meant to be vector and lost its source file, not converting photography. For a flat logo, simple mode is the one to try first, because a mark with 40 paths is editable and a mark with 4,000 is not.

Because there is no public data on what these files look like inside, the first ten files you generate are your evaluation. The checks below are the ones a production designer runs on any supplied vector, AI-made or not, and they take a few minutes per file. They are the vector equivalent of the cutout checks in our transparent-background guide, where the file that looks right in a browser fails on a dark placement.

Ideogram is not the first to generate vectors: Recraft, for one, offers dedicated vector models through its API. But most image models teams already use produce pixels, so the practical comparison is with the two-step workflow they force. Canva announced its own in-house image model, Craft, on the same day, and it is a raster model with a planning stage; Bria's FIBO 2 family, covered later in this series, is raster with structured JSON controls. Each of those is the better tool for a photographic ad or a product shot. None produces a file a sign maker can cut.

So the routing is by deliverable. Anything that will be printed at unknown size, cut, embroidered, engraved or recoloured per campaign is a vector job and now has a direct model. Anything photographic stays raster, and our image API pricing comparison covers that market's per-image prices. Where a brand's colours and type already exist as tokens, the brief to either kind of model should start from them, which is the discipline in our design-system extraction guide.

For web teams specifically, a native SVG has one more property worth planning for: it can be inlined, styled with CSS and animated. A generated illustration whose shapes carry sensible structure can become a themed, dark-mode-aware component rather than an image tag. Whether Vector 1.0's output has that structure is, again, a ten-file test, and it is one our web development team would run before committing a design system to it.