Ideogram Vector 1.0 vs MAI Image 2.6 Preview: 16,000 Votes Against None

MAI Image 2.6 Preview has been through 16,591 blind human comparisons on Artificial Analysis's image-editing board, where it currently sits third at 1,153 Elo. Ideogram Vector 1.0, released by Ideogram on October 22, 2026, has been through none. Not a few, not a vendor-run study — none. That asymmetry, rather than price or feature count, is the actual state of this comparison, and any page that hands you a confident winner is inventing evidence it does not have.

This matters more than it usually would because the two models are being sold against the same brief. Both are pitched at designers. Both claim precise text. Both are positioned as the tool for commercial graphics. One of them has a public scoreboard, and the other one cannot have one, because no scoreboard in the industry measures the thing it produces.

The evidence inventory

Start with what can be checked.

• MAI Image 2.6 Preview (styled MAI-Image-2.6 by Microsoft) — on Artificial Analysis's AA-Image-Editing v2.1 board, read October 23, 2026, it is ranked third at 1,153 Elo with a ±12 interval and 16,591 samples, listed at $38.90 per thousand images. On AA's text-to-image board it sits at 1,165 Elo, rank six.

• MAI's vendor claims — Microsoft says the model is second on Arena's text-to-image board and second on AA's text-to-image board, first on editing, with a 79-Elo gain over MAI Image 2.5 and 91 Elo on text rendering. It also claims the best price-per-Elo in the world. Some of these claims are older than the boards: on AA's editing board today the model is third, not first, and its text-to-image rank is six, not two. Vendor claims about leaderboard position age quickly, which is why the board itself is the better citation.

• Ideogram Vector 1.0 — no independent samples, no benchmark table, no technical report, no public evaluation set. The vendor claim is that outputs are clean, editable SVGs with smooth gradients and sharp typography, delivered with separate layers. Partner rate cards confirm the price; nothing confirms the quality.

• The missing board — there is no vector-image leaderboard anywhere. Recraft, the closest thing to an incumbent in text-to-vector, is not on AA's image boards either, and neither is any other SVG model.

What a blind preference vote actually measures

AA's editing board is Elo from pairwise human votes: two outputs, one prompt, a person picks the better one. That produces a genuinely useful signal about how an image looks and how well it followed the instruction. It is the best public evidence in this matchup, and it is worth taking seriously.

It is also blind to everything that makes a vector file different. A voter comparing two PNGs cannot see whether the winner's shapes are separate objects, whether the gradient is a real gradient definition or a banded fill, whether the lettering is live text or converted outlines, or whether the file will survive being scaled to a billboard. Those are the properties that decide whether an SVG is usable, and none of them are visible in a 512-pixel thumbnail.

So the honest reading of MAI's 1,153 Elo is narrower than it looks: it is strong evidence that MAI Image 2.6 Preview produces preferred raster images. It is zero evidence about whether Ideogram Vector 1.0 produces clean vectors, and zero evidence that it does not.

Why nobody scores SVG

The absence of a vector leaderboard is not an oversight. It is a measurement problem. Raster preference is easy to crowd-source because the output is a picture that anyone can look at. A vector file is a structured document, and its quality is at least four-dimensional: geometric cleanliness, layer structure, colour fidelity, and text editability. Two files that render identically on screen can differ by a factor of fifty in anchor count, and the one with fewer anchors is the one a print shop wants.

No one has built a public prompt set with reference SVGs, and no one has agreed on how to weight anchor count against visual fidelity. Until that exists, every claim about vector quality — Ideogram's included — is a claim about how a file looks, not how it is built.

A five-minute inspection protocol

The useful response to a missing benchmark is to build a small one yourself. Open the SVG in a vector editor and check these seven things. They take less time than reading a leaderboard, and they answer the question the leaderboard cannot.

• Layer count and structure — how many top-level groups does a simple mark produce? A well-structured file has a handful; a traced one has dozens of ungrouped fragments. Ideogram's vectorizer documents a grouping option and states that grouping is not yet available, so check whether the generator does better than the vectorizer here.

• Anchor count on a simple shape — a clean circle is four anchor points. If a circle arrives with two hundred, the model is tracing rather than constructing, and every subsequent edit will fight the geometry.

• Text representation — inspect the file source and see whether lettering is a <text> element or a <path>. This single fact decides whether a typo is a ten-second fix or a re-set in a matching font. No public source confirms which one Ideogram Vector 1.0 emits, and the marketing language does not distinguish between "editable layers" and "editable text".

• Gradient fidelity — are gradients defined as gradient objects, or baked into hundreds of banded fills? The former scales; the latter shows banding the moment the file is enlarged.

• Recolour test — select one element, change its hex value, and confirm that only that element changes. If the mark was flattened into one path, this fails immediately.

• Scale test — render at 10,000 pixels on the long edge and look for rounding artefacts on curves. This is where a native vector model should be flawless and a traced raster will not be.

• File size — a simple mark should be kilobytes. A one-megabyte SVG for a three-shape logo is a traced photograph wearing a vector file's clothes.

Run those seven checks on three outputs and you will know more about Ideogram Vector 1.0 than any published review currently does.

What MAI Image 2.6 Preview is actually built for

Microsoft's model is a diffusion model with flow-matching loss, aimed at high-quality text-to-image plus targeted image-to-image editing: swap an object, change a layout, update text in a scene, clean up an artefact, iterate on a character. It accepts up to five reference images, which is what makes multi-subject composition possible. It takes a 32,000-token context. It has web grounding, which pulls live information for current-events imagery — a genuinely unusual capability for an image model and the reason it can render a plausible weather map or a news graphic with real numbers in it. It is a public preview model sold through Microsoft Foundry, and Microsoft announced its Arena position on September 4, 2026.

Its constraints are equally concrete:

• Output — PNG only, with dynamic aspect ratios and a maximum around 1.5K (roughly 1536×1536). Microsoft's own materials and some Azure documentation disagree on the exact pixel cap, listing either 1,048,576 or 2,359,296 total pixels depending on the endpoint, so verify the limit on the endpoint you actually call.

• Price — $5 per million text input tokens, $8 per million image input tokens and $38 per million image output tokens, which works out to roughly $0.04 per image. The Flash variant is $1.75, $2.50 and $19, or about $0.0195 per image.

• Distribution — sold directly through Microsoft Foundry and the MAI Playground. It is a preview, not a general-availability model, and preview terms apply.

None of that helps with a vector deliverable. All of it is good for mockups, campaign visuals, product edits and anything where the input is a photograph and the output is a prettier photograph.

What the vector model does that no vote can capture

Print production is the obvious answer, and it is the right one: embroidery, vinyl, screen print, etching and signage all consume paths. But the less obvious answer is file economics. A mark that lives in a design system has to be recolourable by token, not by pixel edit. A logo package delivered as SVG is kilobytes; the same mark as a 4K PNG is megabytes, and the PNG still is not the source of truth.

What the vector model cannot do is equally clear. Ideogram Vector 1.0 exposes aspect ratio and seed and not much else in the way of composition control. There are no bounding boxes, no per-element placement, no reference-image count published in the API reference. Complex layouts — a poster with a headline, a subhead, three columns and a logo lockup — are not what this model is documented to do. MAI Image 2.6 Preview, with five references and web grounding, is built for exactly that kind of scene assembly.

Settling it yourself

Neither model is on OrcaRouter's catalogue. That is worth saying plainly, because the practical advice that follows depends on it: for the raster half of this decision you have real alternatives, and for the vector half you effectively have one vendor and a handful of smaller ones.

The way to settle a quality question without a benchmark is to run the same twenty prompts through several models and compare the outputs yourself. That is the workflow OrcaRouter is built for — one API for more than 200 models, provider list prices passed through at 0% markup, and automatic failover so a long evaluation batch does not die because one provider rate-limits halfway through. Neither MAI Image 2.6 Preview nor Ideogram Vector 1.0 is in that catalogue today; when they are, they will be there at list price, not at a marked-up one.

For now the protocol above is the honest substitute. It is also more decision-relevant than a 16,000-vote score, because it measures the file you are actually going to ship.

The verdict that is not a winner

If the output is a raster image and you want the safest quality bet, MAI Image 2.6 Preview is the one with evidence behind it: 16,591 blind comparisons, a top-three editing rank, a published price near four cents an image, and capabilities — five references, web grounding, surgical editing — that the vector model does not have.

If the output is an SVG, the comparison is not close in the other direction, because MAI cannot produce one at any price. The question is not whether Vector 1.0 beats MAI at vectors; it is whether Vector 1.0's vectors are good enough to ship, and that is a question only an inspection can answer.

What you can say with confidence today is narrower than a headline: the only model in this pair with independent evidence is the raster one, the only model that produces the file format print workflows need is the vector one, and the five-minute inspection is the cheapest way to close the gap between those two facts.