MAI Image 2.7 Preview vs Ideogram 4.5: Two Kinds of "Precise," Only One With a Price

Both models were sold on precision, and the word means different things in each pitch. Ideogram 4.5, released September 30, 2026, means pixel-level precision: a mask that says which pixels may change, an edit that runs at the source image's own resolution rather than downscaling it first, and a documented promise that surrounding pixels stay where they were. MAI Image 2.7 Preview, released by Microsoft AI four weeks later on October 27, 2026, means judgment-level precision: restyle a whole scene without breaking it, relight it correctly, rewrite the text so it belongs to the surface instead of floating above it. Neither of these two can be called through a gateway, and neither vendor has said when that changes. See the image models that are routable today. The models you can call are listed with their provider pricing.

Those are not competing claims about the same capability. One is a control surface, the other is a taste model. And the practical difference between them shows up in the least glamorous part of the spec sheet: Ideogram 4.5 publishes a mask convention and a price per image; MAI Image 2.7 Preview publishes neither.

What "Precise" Means When There Is a Mask

A mask is a contract. Ideogram 4.5 takes one, with black marking the region to edit and white marking the region to preserve, alongside up to five input images — one source plus four references. That interface turns an edit into something you can specify rather than something you describe and hope for.

It also makes the model testable in a way a prompt-only editor is not. If a client says "change the bottle label and nothing else," the mask version can be checked mechanically: diff the output against the input outside the masked region and count changed pixels. The prompt version requires a human to look at both images and form an opinion. At five assets that difference is nothing. At five hundred it is the difference between a QA step and a QA department.

Microsoft has published no mask semantics, no bounding-box interface and no coordinate control for MAI Image 2.7 Preview. The model accepts an instruction and a source image. What it changes is decided by the model, and the company's description of the design goals — whole-scene restyling, relighting, restoring an asset, editing text and symbols in context — is a description of edits that are intentionally global. That is a legitimate design choice for the jobs 2.7 is aimed at. It is also the reason you cannot hand it a locked-down regional edit and expect the contract to hold.

Twenty-Four Megapixels In, Twenty-Four Megapixels Out

Ideogram's headline demonstration for 4.5 was an edit on a 4,016×6,016 source — about 24.2 megapixels — performed without downscaling the input first. That is a specific and unusually honest benchmark, because downscaling is where most editing models quietly lose the thing you cared about. An image that gets resampled to two megapixels, edited, and upscaled back has been through two lossy transforms before you see it, and the "preserved" region is not preserved at all; it is a reconstruction.

Native output for 4.5 is 2K, with a 2,048-pixel long edge at the top quality setting and thirteen aspect ratios from 1:1 through 3:1. So the pitch is not that it returns 24 megapixels. It is that a large source survives the round trip without being flattened on the way in.

For MAI Image 2.7 Preview, no resolution has been published at all. The predecessor MAI-Image-2.6 capped output near 1536×1536, and assuming 2.7 matches would be inference rather than reporting. When one side of a comparison has a published number and the other has silence, the published number is not automatically the better one — but it is the only one you can engineer against.

Typography Is Where the Two Pitches Collide

Text rendering is the one capability both models claim outright, and it is worth being specific about why it is hard in both directions.

Ideogram's reputation was built on typography in generation — legible words, correct spelling, layouts that respect a headline's hierarchy. In 4.5 that strength carries into edits, and the model also keeps a text-to-image mode, so a team can generate a layout with real words in it and then revise the same layout without changing tools.

Microsoft's claim for 2.7 is narrower and, in one respect, harder: changing existing text and symbols while keeping them integrated — matching the perspective of a sign, the curvature of a label, the lighting on a printed surface. That is a different problem from rendering a headline cleanly on a flat canvas, and it is the problem that shows up when a campaign needs the same packaging shot in six languages. Artificial Analysis has 2.7 first on text and symbol editing among the seven actions it evaluates, which is the strongest independent evidence either model has on this specific axis.

The catch is that the text result is a category rank, not a per-image guarantee, and it says nothing about how many attempts a hard case takes. Microsoft says 2.7 improved across all ten use cases and all seven editing actions relative to 2.6, with a 24-point lead in scene and style; those are the company's own cuts of an independent board, and the underlying runs are not public.

The Price Ladder, and the Model That Has None

Ideogram 4.5 prices by quality tier rather than resolution, which is an unusual and mostly sensible choice: you pay for how hard the model thinks, not for how many pixels come back. The published tiers run roughly $0.03 at Low, $0.06 at Medium and $0.22 at High per image, with an additional very-low tier available only for edits, quoted by some resellers from about $0.008. Prompts expand through a Magic Prompt toggle, seeds are exposed, and a single request can return up to eight images — which means the effective cost per usable output depends heavily on how many candidates you generate before one is right.

That last point deserves emphasis because it inverts the headline price. A model at $0.03 that needs four attempts costs twelve cents; a model at $0.22 that lands on the first try is the cheaper option. The tier list is a menu, not a bill.

MAI Image 2.7 Preview has no price because it has no commercial endpoint. It is available in MAI Playground, on Arena, and on Artificial Analysis. There is no API, no Foundry listing, and no rate card. The closest thing to a forecast is the predecessor's Foundry pricing — $5 per million text input tokens, $8 per million image input tokens, $38 per million image output tokens, roughly four cents an image — which tells you roughly where Microsoft prices image work and nothing about where 2.7 will land.

Where the Scores Sit, and Why the Gap Is Smaller Than It Looks

MAI Image 2.7 Preview is second on AA-Image-Editing v2.1 with 1,193 Elo, six behind GPT Image 2.5 Sunburst. Ideogram 4.5 was reported at 1,064 Elo and twenty-third on the editing board, with 1,014 Elo and thirty-fourth of 167 on the text-to-image board.

Those two editing figures are not a clean subtraction, for the same reason it is never a clean subtraction across board revisions: the 4.5 number comes from its own review window, and the 2.7 number from an October 22 snapshot of a board that has been rebuilt at least once this year. The direction of the gap is not in doubt — a top-three model against a twenty-third-place model is a real separation — but the size of it is not a measurement.

What the placement does say is that Ideogram's precision pitch has not translated into a frontier editing score. Its text-to-image score, 1,014, sits barely above the previous generation's 1,011. The company's argument has always been about control and typography rather than leaderboard position, and 4.5 is consistent with that: a controlled tool at a mid-table rank, priced per tier, with a mask interface that makes it auditable.

The Open-Weights Promise

Ideogram said at launch that 4.5's weights would be released soon. They have not shipped. That matters because it is the single feature that would make the model categorically different from anything Microsoft is offering — a downloadable checkpoint is an architecture decision, a hosted API is a dependency.

Until the weights appear, the claim is a roadmap item and should be treated as one. The model is available today through Ideogram's own app and API and through a long list of partner platforms, which is a real form of availability but not the same as portability.

Choosing Between Them

Pick Ideogram 4.5 if the edit has to be bounded. If you can draw the region that may change, if the source is large and must survive at its own resolution, if you need a text-to-image mode and an edit mode in one model, and if you need to know the cost before you start — 4.5 answers all four questions in its documentation. It is also the only one of the two you can put behind an API today.

Pick MAI Image 2.7 Preview if the edit is judgment-heavy and the volume is low. Restyling a scene, relighting a photograph, rewriting a sign in a way that respects the surface — these are the tasks where a 24-point category lead is worth something, and Playground is a reasonable place to do twenty of them by hand.

What you should not do is treat the leaderboard gap as a purchase recommendation. The model in second place cannot be called, and the model in twenty-third place can be called for six cents at Medium. For a team deciding what to build on this quarter, that is the whole comparison. OrcaRouter does not host either of them, and neither vendor has announced plans to change that; where a gateway is useful here is in the models you would fall back to when a preview-grade editor is unavailable, since the OpenAI tiers bracketing 2.7 on that same board are routable on one key at list price and can absorb a failed call without a code change.