Ideogram Vector 1.0: What the SVG File Actually Contains

A screen-print shop quotes four hundred dollars to redraw a customer's JPEG logo before it can go on a press. That quote, not the price of a generation, is what Ideogram Vector 1.0 is competing with. Ideogram released the model on October 22, 2026 — its first that outputs vector files rather than pixels — and it does two things: it draws an SVG from a text prompt, and it converts an uploaded raster image into an SVG. The pitch is that a designer opens the result and keeps working, instead of opening a JPEG and starting over.

Whether that pitch survives contact with a real file is the entire question, and it is not a question any image leaderboard can answer. So this is a piece about the file: what is in it, what the API says about it, and what it costs to find out. Every quality claim below is Ideogram's; there are no independent measurements of SVG output quality from anyone yet, including us.

Two endpoints, and they are not the same product

The API is documented as POST /v2/svg/generate/ideogram-vector-1 for prompt-to-vector and POST /v2/svg/vectorize/ideogram-vector-1 for image-to-vector. Both take multipart form data, both accept a dry_run flag, and both return a generation_id you poll for the finished asset.

The split matters because the two paths have different quality ceilings. Generation builds the artwork from scratch, so the model decides how many shapes to use and how to organise them. Vectorisation has to infer geometry from an existing picture, which is a harder and lossier problem. Ideogram gives the vectorizer two modes — default, which "preserves more fine detail," and simple, which "draws fewer, simpler shapes." That is a dial between fidelity and anchor-point count, and there is no mode that gives you both.

The generate endpoint takes a prompt of 1 to 9,500 characters plus optional image references, either as uploaded files or as image_asset_identifiers pointing at assets already in your account. The vectorizer takes a JPEG, PNG or WEBP source of at most 50 MB. Neither endpoint documents a cap on output complexity, which is an odd omission for a format whose whole value is complexity you can edit.

What "editable" has to mean to be worth anything

Every vector model says "editable." The word is doing a lot of work, and it is worth being concrete about which of these properties you actually get, because they are not the same claim:

• Separate objects — can you select the letterform without selecting the background? Ideogram says yes for generated SVG: shapes arrive as separate, individually addressable layers.

• Live colour — are fills and strokes stored as colour values you can change, or flattened into paths? Ideogram says editable colours.

• Preserved gradients — do smooth gradients survive as gradients, or get chopped into banded strips of solid colour? Ideogram claims smooth gradients, which is the harder of the two to deliver and the one to check first.

• Live typography — is text real text, or outlines of text? Almost every generative vector tool ships outlines. Outlines are fine for a logo and useless for a wordmark that marketing will re-set in a different weight.

• Anchor economy — how many points does a simple shape take? A circle drawn with four points is editable; a circle drawn with ninety is a file nobody wants to touch.

Only the first three are claims Ideogram makes explicitly. The last two are the ones a designer checks in the first thirty seconds, and they are the ones no vendor page can settle for you. Open the file, select a shape, and count the points.

The one line in the reference docs that changes a workflow

On the vectorize endpoint, the parameter list includes a boolean called group, described as "Request grouping of shapes into layers." Immediately after it, the docs say: "Grouping is not yet available."

That is a small sentence with a large consequence. It means prompt-generated SVG arrives layered, and image-converted SVG arrives as a flat pile of paths. If your workflow is "scan the client's old logo, clean it up, hand back a layered file," the second half of that sentence is the part you still do by hand. The vectorizer gets you paths; it does not get you structure.

It is also the most honest thing on the page. A vendor shipping a feature flag for a capability it has not built yet is telling you the roadmap in public, and it is worth reading the vectorize docs as a description of what is coming rather than what is here. If grouping lands, the case for vectorising an archive of old assets gets much stronger. Until it does, budget for a grouping pass.

Pricing that is built for quoting

Ideogram publishes per-file rates rather than per-token ones, which for this product category is the right unit: you are buying an asset, not a quantity of computation.

• Low quality — $0.13 per generated SVG

• Medium quality — $0.16 per generated SVG

• High quality — $0.20 per generated SVG

• High quality with reference images — $0.32 per generated SVG

• Image to vector — $0.10 per conversion, in either mode

Those are Ideogram's published API rates, vendor-reported and unaudited by us. What makes them usable rather than merely published is dry_run: set it true and the request is validated and priced but not generated, not stored and not billed, returning a PriceQuote object instead. That turns cost estimation into a loop you can run over a hundred prompts before spending anything, and it is the single most practical thing in the release.

What is not published is a batch tier. There is no discounted asynchronous lane the way some token-billed image APIs offer one. Volume is governed by your account's concurrent-generation limit, and when you exceed it the API returns 429. Plan throughput around that rather than around a discount that does not exist.

What the model will not do

There is no edit endpoint. You cannot ask Ideogram Vector 1.0 to change the background of an SVG it just made, swap a colour, or nudge a shape two units left. You can generate a new file, or vectorise an existing image. That is the whole surface. If your work is iterative — twenty variants, small adjustments, a client who wants the mark slightly bigger — the iteration has to happen in your design tool, not in the model.

It also will not tell you whether the output is good. There is no published complexity limit, no path-count guidance, and no independent benchmark, because none of the image arenas measure vector output. A pairwise preference test on flattened previews would rank a beautiful raster trace above a clean file with forty anchor points. The instrument does not exist yet.

Where this lands in a real production line

The jobs that change are the ones where the deliverable is a file a machine will cut, etch, embroider or print at arbitrary size: signage, apparel, packaging dielines, brand marks, icon sets. In those jobs the alternative to a generated SVG is not a different model — it is a human with Illustrator and an hourly rate, and the comparison is $0.20 against $80.

The jobs that do not change are photographic, scene-based, or iterative. A product shot with convincing light is not a vector problem, and neither is a campaign image that will be edited nine times before it ships.

Ideogram Vector 1.0 is not on OrcaRouter's catalogue; it is available through Ideogram's own app and API and several third-party platforms. What we do carry is the raster side of the same workflow — if you are generating exploratory images first and vectorising the winner afterwards, that exploratory half can run through one API with automatic failover, so a slow provider does not stall the batch while the vector call is a separate, deliberate step you make once.

The test worth running before you build on it

Generate ten files across the range you actually need — a wordmark, a flat icon, a gradient-heavy illustration, a dense label — and for each one write down four numbers: anchor points, whether type is live or outlined, whether gradients banded, and whether the object count matches what you asked for. Then vectorise the same source images and do it again. That is maybe an hour of work, and it produces the only evidence about this model that currently exists outside Ideogram's own marketing.

The honest summary is that Ideogram Vector 1.0 has shipped a real capability with a documented seam in it. The generation path looks like the finished product. The vectorisation path is a beta wearing a 1.0 label, and the gap between them is exactly the gap between "we can make a file" and "we can make a file from yours."