Ideogram Vector 1.0 vs MAI Image 2.6 Preview: One Ships, One Is Still a Preview

The most important difference between Ideogram Vector 1.0 and MAI Image 2.6 Preview is not in either model's output. It is in the second word of Microsoft's model name. Ideogram shipped Vector 1.0 as a generally available product on October 22, 2026, with a published per-file price and an API reference. MAI Image 2.6 Preview has been sitting in a gated public preview on Microsoft Foundry since roughly the start of September 2026, carries no service-level agreement, and Microsoft's own guidance does not recommend it for production. One of these is a purchase. The other is a trial.

That asymmetry does not make the comparison boring. It makes it a genuinely useful one, because the two models are the best available answer to opposite questions: what do I generate at scale, and what do I owe the person who has to deploy it?

What "Preview" removes before you generate anything

Preview status is not a vibe. It is a set of concrete absences that a team has to plan around.

• No SLA. Microsoft's Foundry documentation states plainly that the preview is not recommended for production workloads. If the endpoint is slow or returns nothing on a Tuesday, you have no contractual remedy.

• Rate limits you will hit. The MAI Playground is reported at roughly two requests per minute and twenty images per day. That is a demo budget, not a pipeline.

• No published price. Microsoft has not issued an official rate card for 2.6 Preview. Third-party trackers list figures in the region of $38.90 per thousand images, which is a reasonable estimate and is not a quote.

• No GA date announced. Any roadmap you build is an assumption.

Ideogram Vector 1.0 has the opposite profile. Its price is published per tier, its endpoints are documented down to the query parameter, and its dry_run flag will quote a request without billing it. For a team that needs a number in a spreadsheet before the quarter closes, that is not a minor convenience.

Where MAI Image 2.6 Preview is genuinely ahead

None of the above says Microsoft built a weaker model. On the independent numbers it is clearly the stronger general image model, by a wide margin.

On Artificial Analysis's AA-Image-Editing v2.1 leaderboard, MAI-Image-2.6 sits third at 1,153 Elo with 16,591 samples, at a listed $38.90 per thousand images. The faster MAI-Image-2.6-Flash is sixth at 1,144 Elo across 10,701 samples at $19.50. Ideogram Vector 1.0 is not on that leaderboard at all — not because it scored badly, but because the board measures instruction-based editing and Vector 1.0 does not perform instruction-based editing.

What Microsoft's model does that Ideogram's cannot:

• Instruction editing — change one thing in an image you already have, at leaderboard-leading quality, with multi-turn refinement.

• Multi-reference composition — up to five reference images for consistency across a series, which is the feature that makes character and product work possible.

• Web grounding — a web_grounding control that pulls in Bing Search context for real-world entities. Ask for a specific stadium and it can look up what that stadium looks like.

• Automatic framing — auto_aspect_ratio lets the model choose the crop rather than forcing you to specify one.

• Prompt headroom — up to 32,000 tokens per the Foundry guide, which is far more instruction than a vector prompt can hold. Ideogram Vector 1.0 caps at 9,500 characters.

Where the vector model is ahead, and it is not a small lead

Ideogram Vector 1.0 produces a different kind of object. The output is SVG, not PNG, and that changes what you can do after generation.

• Editable geometry — shapes, gradients and type arrive as objects rather than pixels. Ideogram's launch material claims elements are individually movable and recolourable; that is a vendor claim and, as of this writing, no independent workflow test has been published.

• Text that stays text — in the best case, a wordmark in the output is a live text object, not outlines. That is the difference between a one-minute legal edit and a redraw.

• Resolution independence — an SVG re-renders at any size. MAI Image 2.6 Preview caps at around 1.5K per side, roughly 2.36 megapixels total, with a 768-pixel minimum. For anything destined for print, that ceiling is the whole argument.

• A conversion path — the vectorize endpoint takes a JPEG, PNG or WEBP up to 50 MB and returns SVG, which means you can run the exploration in a raster model and convert the winner. Read the caveat before you build on it: the group parameter is documented as "request grouping of shapes into layers. Grouping is not yet available," so a converted flat image comes back as geometry without layer structure.

The 1.5K ceiling deserves its own paragraph

Microsoft's own model card puts the maximum at about 2,359,296 total pixels, which is a square a little under 1,536 pixels on a side. That is fine for editorial and social. It is not fine for a poster, and it is nowhere near a packaging dieline. Every raster model in this class has the same problem, which is why the workaround industry — upscalers, then a manual trace — exists and why it costs what it costs.

Ideogram Vector 1.0 sidesteps the ceiling rather than beating it, because the output has no native resolution. Whether the SVG is genuinely clean at large scale is a separate question from whether it is infinitely scalable, and only the first one is worth arguing about. A traced PNG scaled to a billboard is still a traced PNG; it just has more vertices.

Questions your procurement form will ask

Can we put MAI Image 2.6 Preview in a customer-facing product?

You can, and Microsoft's documentation advises against it. The absence of an SLA and of a published price means you are accepting availability and cost risk that you cannot put in a contract. The usual compromise is to keep it behind a feature flag with a fallback model on the same key, so a preview outage degrades one path rather than the product.

Is Ideogram Vector 1.0 generally available, or is it also early?

It is GA as of October 22, 2026, with published pricing and public API documentation. GA is a statement about the commercial relationship, not about maturity — the model is one day old at the time of writing, the layer claim is untested independently, and the vectorize endpoint ships with a documented gap. Treat it as a shipped product with an unproven feature set, which is a different risk profile from a preview, not an absence of risk.

Can we get both on one contract?

Not currently. Ideogram's vector endpoints are available through the vendor's own API and several third-party platforms, and OrcaRouter does not route them. What we do provide is a single key for the raster side of the same workflow — GPT Image 2 and Gemini 3.1 Flash Image among 200-plus models, at provider list price with 0% markup — so a preview model can be swapped for a stable one without a second procurement cycle.

The technical gotcha that decides architecture

MAI Image 2.6 Preview is an edit-first model. Ideogram Vector 1.0 has no edit endpoint at all. That means a pipeline built around "generate a draft, then refine it" cannot simply substitute one for the other at the last step, because there is no refine step on the vector side. The realistic shape is two stages: draft and refine in a raster model with reference control, then convert the approved composition to SVG.

If that is the plan, test the conversion early. Generate a mark in MAI Image 2.6 Preview, convert it through the vectorize endpoint at mode: default, and open the result in Illustrator. Count the objects, check whether the letterforms are text or outlines, and try changing one fill colour. Fifteen minutes of this will tell you more about whether the workflow holds than any launch coverage can, and it costs about ten cents a conversion.

Who should pick which

Pick MAI Image 2.6 Preview when the deliverable is a raster image that needs to look right: marketing imagery, product scenes, editorial illustration, anything with people or real-world grounding. Accept the preview terms deliberately, or wait for GA and use the Flash variant's cheaper rate in the meantime.

Pick Ideogram Vector 1.0 when the deliverable is a file that will be resized, recoloured or printed, and when you can tolerate a one-shot generation with no regional correction. Budget for the possibility that the layer structure is less clean than advertised, and verify it on a sample before committing a brand system to it.

What you should not do is treat this as a quality contest. On general image quality, Microsoft wins on the evidence that exists. On the question of whether the thing you made can leave the building, Ideogram is the only one of the two that answers yes.