Ideogram 4.5 vs MAI-Image-2.5-Pro: A New Editor Meets a Superseded One

There is an awkward fact buried in this matchup that neither vendor's marketing surfaces. MAI-Image-2.5-Pro is the opponent named in the comparison, and it is genuinely strong — ninth in the world on Artificial Analysis's editing board, ninth on the same board where you can also see that Microsoft has since shipped MAI-Image-2.6 and MAI-Image-2.6-Flash, both of which outscore it and cost a fraction as much. Meanwhile Ideogram 4.5, live since September 30, 2026, is the newest editor on that board and ranks 23rd. So the useful version of this article is not "which of these two wins." It is: a new specialist with an unmeasured claim against a well-tested incumbent whose own vendor has moved past it.

The direct answer between the two named models: MAI-Image-2.5-Pro is ahead by 44 Elo on editing and roughly three times cheaper per image, on a sample five times larger. Everything that follows is about why that is not the end of the story.

Both models in one short paragraph each

• Ideogram 4.5 — shipped September 30, 2026. A precise-edit model with two operations, generate and edit, at four rendering speeds. Its claim is fidelity across sequences: edits preserve the boundary of the changed region so the result can be stitched back into a source up to 4,016 × 6,016 without re-rendering the rest of the frame. Text-to-image 34th of 167 at 1,014 Elo; image editing 23rd at 1,064 Elo; $0.03 to $0.22 per image by speed setting.

• MAI-Image-2.5-Pro — Microsoft's quality-tier image model, in public preview on Azure AI Foundry since July 23, 2026. A focused, surgical editor: targeted object replacement, layout changes, in-image text updates, artifact cleanup, with Microsoft claiming 94% multi-image character consistency and 96.8% text-rendering accuracy across five languages — both vendor-reported and unverified. Text-to-image 12th at 1,100 Elo; image editing 9th at 1,108 Elo; $108.50 per 1,000 images on the board's conversion.

Both are closed, hosted and metered. Neither is a general-purpose visual model. Both are aimed at the same buyer: someone whose images already exist and need to change without being rebuilt.

The board, and the row underneath the row

Artificial Analysis's editing board is blind pairwise human preference. Take the two rows relevant here:

• Editing — MAI-Image-2.5-Pro 9th at 1,108 ± 8 Elo, 14,733 votes, $108.50 per 1,000. Ideogram 4.5 23rd at 1,064 ± 11 Elo, 2,382 votes, $220.00 per 1,000.

• Text-to-image — MAI-Image-2.5-Pro 12th at 1,100 ± 8 Elo, 14,523 votes. Ideogram 4.5 34th at 1,014 ± 11 Elo, 2,247 votes.

Now the rows Microsoft would rather you did not read next to them, because they change what "incumbent" means:

• MAI-Image-2.6 (August 2026) — editing 3rd at 1,137 Elo; text-to-image 5th at 1,150 Elo; $38.90 per 1,000 images.

• MAI-Image-2.6-Flash (September 2026) — editing 4th at 1,125 Elo; text-to-image 9th at 1,105 Elo; $19.50 per 1,000 images.

• MAI-Image-2.5 (June 2026) — editing 7th at 1,113 Elo, $48.10 per 1,000.

Read those together and MAI-Image-2.5-Pro's position becomes clear, and unflattering in a way that has nothing to do with Ideogram. It is the most expensive Microsoft image model on the board at $108.50, and it is beaten on editing by three of its own siblings, including one that costs five and a half times less. Its July preview pricing was always going to be revisited; what the board shows is that Microsoft has since built cheaper models that score higher, which is the ordinary shape of a product line maturing and also the reason a buyer comparing "Ideogram 4.5 vs MAI-Image-2.5-Pro" should widen the frame by one row.

Two different editing philosophies

The Elo gap is 44 points; the design gap is wider.

Microsoft's pitch for MAI-Image-2.5-Pro is consistency. The model is built so that you change one thing and nothing else moves — subject identity, lighting and texture are held stable across iterations, which is what the 94% multi-image character-consistency claim is about. It is a general-purpose surgical editor: swap an object, fix a layout, correct the type on a sign, clean up an artifact. It is not built around one headline property; it is built to do the ordinary jobs well and consistently.

Ideogram's pitch for 4.5 is narrower and more technical. It is not claiming better-looking edits than anyone else — its own page is careful to demo a behaviour rather than print a rank. It is claiming that edits stop degrading as you stack them, and it demonstrates that at source resolution on a 24-megapixel file. Where Microsoft sells consistency of subject, Ideogram sells consistency of the untouched pixels.

Those are different failure modes. A model can preserve a face perfectly and still re-render the wallpaper. A model can hold the frame bit-identical and still drift on identity. Nobody publishes a benchmark that captures either across a sequence of edits, which is why both vendors are selling claims that sit outside the arena.

Price, at the setting that matters

• Ideogram 4.5 — $0.03 per image at Low, $0.06 at Medium, $0.22 at High. Native 2K, cost tracks rendering speed rather than resolution.

• MAI-Image-2.5-Pro — token-metered at $5 per million text-input tokens, $8 per million image-input, $106 per million image-output, which Artificial Analysis converts to $108.50 per 1,000 images. Preview pricing; a general-availability rate card could move it, and given the sibling rows it is hard to see it holding.

• The wider Microsoft line — $19.50 per 1,000 for MAI-Image-2.6-Flash against MAI-Image-2.5-Pro's $108.50 is a five-and-a-half-times gap for a model that ranks higher on both boards.

So the per-image arithmetic is: Ideogram 4.5 at its High setting ($0.22) costs about twice MAI-Image-2.5-Pro's converted rate, and roughly eleven times MAI-Image-2.6-Flash's. At its Low setting it undercuts everything here. The setting is the whole decision, and the drift demonstration Ideogram is selling lives at the expensive end.

What a sensible buyer actually does here

Three decisions hide in this comparison, and they have different answers.

If the question is "which of these two editors should I use," MAI-Image-2.5-Pro wins on evidence — 44 Elo ahead, on five times the votes, at roughly a third of the price. Ideogram 4.5's counter is a sequence-fidelity claim that no arena tests. That is a real claim and worth testing on your own images, but it is not currently evidence.

If the question is "I have decided on Microsoft, which model," MAI-Image-2.5-Pro is close to the wrong answer. MAI-Image-2.6-Flash ranks fourth in the world at editing, ninth at generation, and costs $19.50 per 1,000 — a fifth of the 2.5-Pro rate. Unless you have a specific reason 2.5-Pro wins on your workload, the newer cheaper sibling is the better default.

If the question is "is the model I use today good enough that I should not switch," the honest test is the one nobody publishes: a ten-turn edit sequence on your own files, diffing the regions you never asked the model to touch, and pricing the run per finished sequence rather than per generated image. Do that before buying either. It takes an afternoon and it answers a question that 44 Elo and a $220 rate card cannot.

Where OrcaRouter fits, and where it does not

Neither model in the headline is on our catalogue. Ideogram does not appear on OrcaRouter at all, and MAI-Image-2.5-Pro is a Microsoft Foundry preview reached through Azure. Saying otherwise would be inaccurate, and the same applies to the newer Microsoft rows — 2.6 and 2.6-Flash are Foundry models, not routes we serve.

What is on the catalogue for image work today is narrower: OpenAI's GPT-Image-1.5 at $8 per million input tokens and $32 per million output, GPT-Image-2 at $8 and $30, xAI's earlier grok/grok-imagine-image at $0.20 per call, and Google's Imagen 4 family. All at provider list price with no markup added, which means a vendor's own price cut lands on our side the same day without a renegotiation — the reason a per-image rate card you read here is the vendor's rate card, not ours.

The practical value in a comparison like this one is for the trial phase rather than the production phase. When a model arrives on the catalogue, the cheap way to test a fidelity claim is to point a slice of traffic at it, keep your incumbent as automatic failover, and let a bad batch fail over rather than fail. On this specific matchup, though, both candidates sit outside the catalogue, and the answer to "where do I get it" is each vendor's own API and the third-party platforms that resell them.

The verdict, with the caveat attached

On the measures that exist, MAI-Image-2.5-Pro beats Ideogram 4.5 at editing, beats it at generation, costs less per image and is backed by five times the voter sample. If you need an answer today and those are the only two names on your list, that is the answer.

Two caveats belong next to it. The first is that the measure does not test the thing Ideogram is selling — multi-turn fidelity is invisible to a single-edit arena, so 23rd place is not a refutation of the drift claim, just as the drift claim is not evidence. The second is that the incumbent is being superseded by its own vendor: at $19.50 per 1,000 and fourth place at editing, MAI-Image-2.6-Flash is the Microsoft image model most buyers should be comparing against, and the fact that this article is framed around a more expensive, lower-scoring sibling says more about how the matchup was drawn than about either lab's engineering.

Watch for two things. An independent multi-turn fidelity test — the first one published settles an argument both vendors are currently making without receipts. And any movement on MAI-Image-2.5-Pro's preview pricing, because a $108.50 rate card sitting above three cheaper, higher-scoring siblings is a position that does not survive contact with a sales cycle.