Ideogram 4.5 vs MAI Image 2.6 Preview: A Specialist Against a Platform

Ideogram 4.5 exists to do one thing: edit an image repeatedly without letting it degrade. MAI Image 2.6 Preview, Microsoft AI's flagship image model, exists to do everything at once — generate, edit with multiple references, ground a prompt in live web content, and pick its own aspect ratio — and it has been callable in public preview through Microsoft Foundry since September 4, 2026. Put them side by side and the numbers point one way and the product shapes point the other. Ideogram wins the thing it was built for; Microsoft wins almost every other axis, including the ones that decide whether a team can actually deploy it.

What each one is

Ideogram 4.5 went live on September 30, 2026 across Ideogram's app, its own API, its MCP integration and its apps collection, plus partner platforms. It generates and edits at four rendering speeds, outputs natively at 2K, and demonstrates localised edits on a 4,016 × 6,016 source — 24.2 megapixels — without downscaling. The vendor's claim is that repeated passes stop drifting: edges stay put, texture survives, an edited crop stitches back into the original. Weights are not published.

MAI Image 2.6 is Microsoft AI's strongest image model and the top of a lineage that runs MAI-Image-1, MAI-Image-2 in March 2026, MAI-Image-2.5 in June, MAI-Image-2.5-Pro in July and MAI-Image-2-Efficient along the way. It debuted on Arena's boards in August 2026 with a No. 2 placement, and on September 4, 2026 Microsoft moved it into public preview in Microsoft Foundry and MAI Playground alongside a new MAI-Image-2.6-Flash for high-throughput workloads. Capabilities on the card are broad: multi-reference editing that combines people, products, styles and scenes from several images, web grounding, dynamic aspect ratios, and output up to 1.5K resolution. It is also closed, and it runs only on Microsoft's cloud.

The line-up

• Access — public app, API and MCP integration since September 30, 2026 vs public preview in Microsoft Foundry and MAI Playground since September 4, 2026.

• Price — $0.03 per image at Low, $0.06 at Medium and $0.22 at High, on a vendor rate card vs no published rate card on Microsoft's news post; pricing sits behind the Foundry model card, and the independent boards list the model at $38.90 per 1,000 images.

• Text-to-image score — rank 34 at 1,013 Elo from 2,278 samples vs 1,161 Elo from 9,410 samples, released August 2026.

• Editing score — rank 23 at 1,064 Elo from 2,459 samples at $220.00 per 1,000 images vs rank 3 at 1,137 Elo from 15,684 samples at $38.90 per 1,000.

• Output — native 2K with edits demonstrated at 4,016 × 6,016 vs up to 1.5K resolution per Microsoft's card.

• Capability shape — one sharp edge (multi-turn edit fidelity) vs a broad set: multi-reference editing, web grounding, dynamic aspect ratios, and a faster Flash sibling.

Same board, same panel, and it is not close

This is the rare matchup where a clean comparison exists, because both models sit on Artificial Analysis with rows measured by the same panel and the same vote pool.

Text-to-image: MAI-Image-2.6 at 1,161 Elo from 9,410 samples against Ideogram 4.5 (High) at 1,013 Elo from 2,278 samples. That is 148 Elo on four times the sample, and the price column reads $38.90 against $100.00 per thousand images.

Image editing: MAI-Image-2.6 at 1,137 Elo from 15,684 samples against Ideogram 4.5 (High) at 1,064 Elo from 2,459 samples, on an independent conversion of $38.90 against $220.00 per thousand.

Microsoft's own September 4 post claimed the No. 2 text-to-image and No. 1 image-editing positions on this board, both flagged "as of Sep 4, 2026". Read a month later, the text-to-image placement still holds and the editing claim no longer does: GPT Image 2.5 Sunburst landed in September and sits first on that board at 1,182 Elo, with MAI-Image-2.6 third at 1,137. That is not a knock on the model — it is what happens to any "No. 1 today" claim in a market that ships weekly, and it is the reason this article quotes a date with every rank.

The difference that matters is not the score

If the boards decided this, it would be over. They don't, for two reasons.

The first is that neither board measures Ideogram's actual claim. Editing arenas show a voter a source image and two single-instruction results; they never show a voter the tenth edit in a sequence against the first, which is the drift that Ideogram built 4.5 to prevent. A 73-point editing gap tells you MAI Image 2.6 is preferred for single edits. It tells you nothing about turn ten on a large file, and turn ten is the thing a catalogue team, a signage swap or a photo restoration actually depends on.

The second is availability shape. Preview is not the same as a product. Foundry public preview is the right home for evaluation and internal tooling, and Microsoft has not attached general-availability terms to it — no availability commitment, no deprecation notice, no throughput guarantee. Microsoft's model is also reachable only through Microsoft's cloud and playground, so there is no second provider to fall back to when a preview endpoint is throttled. Ideogram 4.5 ships on a published ladder with a rate card you can put in a budget, and Ideogram's own API is a commitment rather than a preview.

Then there is the price direction, which is genuinely uncomfortable for the newer model. In the absence of a published Microsoft rate card, the only comparable numbers are the boards' conversions: MAI-Image-2.6 at $38.90 per thousand against Ideogram 4.5 at $100.00 for text-to-image and $220.00 for editing. A model with better scores on both boards and a lower converted price is not a hard sell — it is the default.

Where each one actually earns its place

Take Ideogram 4.5 when the expensive part of your workflow is repairing approved work rather than producing new work: colourway variants for a catalogue, signage translated across a campaign, a damaged photograph restored without repainting what survived. The 24.2-megapixel demonstration and the preserved-edge claim are aimed exactly there, and no competing model is making that claim with comparable evidence. Test it on your worst-case sequence before you commit; the claim is unreproduced by anyone outside the vendor.

Take MAI Image 2.6 when you need breadth and speed of integration. Multi-reference editing with several source images, web grounding and dynamic aspect ratios cover most production briefs without a second model in the pipeline, and MAI-Image-2.6-Flash exists for the volume tier — Microsoft's own figure is 2.8× faster than GPT-Image-2-Medium, measured in an internal load test at 100 requests per minute at 1024 × 1024, which is vendor-reported and unreproduced. If your team already lives in Azure, the integration cost is near zero and that may outweigh 24 megapixels of edit fidelity.

Take both if you are running a real content operation. A platform model for generation and volume, a specialist for the passes that have to be pixel-faithful, and a rule about which one owns which step. The mistake is treating this as a single purchase.

Routing, when the two endpoints have different guarantees

Nothing in this matchup is on OrcaRouter's catalogue. Ideogram 4.5 is available through the vendor's own API and several third-party platforms; MAI Image 2.6 Preview is reachable only through Microsoft Foundry and MAI Playground. Our image line sits elsewhere — G​oogle's Gemini 3.1 Flash Image and Imagen 4 identifiers, O​penAI's GPT-Image models — and saying we route either of these would not survive a single check.

What a routing layer is genuinely for here is the asymmetry in those two availability stories. A preview endpoint with no SLA is exactly the case where automatic failover earns its keep: the call lands on a served alternative instead of throwing at 2 a.m., and you find out in the logs rather than in a support ticket. Per-call pricing at provider list price with nothing added on top means a vendor price cut or a preview's first published rate appears on your invoice the day it lands — which matters when the comparison you are running is 2.6× on the editing line. And if you want to A/B a preview against a specialist without two contracts, one key is the cheapest way to do it.

What to do now

If you are choosing one model today, the honest answer is that the better model is also the cheaper one and the one you can only get in preview. Pick MAI Image 2.6 if you can live inside Microsoft's cloud and preview terms; pick Ideogram 4.5 if you need a rate card you can forecast, a first-party API, or the specific edit-fidelity behaviour, and accept that you are paying roughly two and a half times the converted editing rate for it.

Two things would settle it. The first is a multi-turn fidelity benchmark: same image, ten sequential edits, measured divergence. Until somebody runs one, the property that makes Ideogram 4.5 interesting is unmeasured and the property that makes MAI Image 2.6 interesting is measured. The second is Microsoft moving MAI Image 2.6 from preview to general availability with a published rate card; that removes the last practical objection to the platform play and makes the specialist argument rest entirely on drift.