Title card comparing MAI Image 2.7 Preview with MAI Image 2.7 Flash: the Preview at 1,193 Elo on 27 October 2026 with its row withdrawn from the board, against Flash at 1,179 Elo on 17 November 2026 at about $10 per 1,000 images, with a bracket between them reading 14 points, inside the error bars.
Guides & Insights

MAI Image 2.7 Flash vs MAI Image 2.7 Preview: The Prototype Was Level

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Microsoft's October prototype and its November budget model land within noise of each other on the same editing board, and only one of the two can be bought. MAI Image 2.7 Preview put up 1,193 Elo on Artificial Analysis's AA-Image-Editing v2.1 leaderboard on 27 October 2026 — No. 2 in the world, from 4,891 votes, with no API, no price and no text-to-image mode. MAI Image 2.7 Flash, released on 17 November 2026 alongside the flagship it is derived from, sits at 1,179 Elo and No. 7 — fourteen points lower, comfortably inside the error bars, and available now in public preview through Microsoft Foundry at roughly $10 per thousand images.

So the comparison is not "which of these is better." It is that the model you can actually call scores the same as the demo you could only try in a playground, and does two extra jobs for a price Microsoft finally published. The interesting question is how much weight those fourteen points can carry. Less than the ranking column suggests, and the reason why is worth walking through before anyone quotes either number in a deck.

Two rows, one board, one of them retired

Artificial Analysis rebuilt its image editing board as v2.1 in October, and both models were scored on that same version — which is what makes the comparison legitimate at all. The Preview entered it at No. 2 on 27 October, six points behind GPT Image 2.5 Sunburst (max) at 1,199. Flash entered it in November at No. 7, in a band spanning ranks five to seven, with 5,891 votes.

The Preview's row is no longer on the board. Flash's is. That is the first asymmetry, and it is not cosmetic: a withdrawn row cannot accumulate votes, so the two figures describe a fixed October snapshot against a November reading that will keep moving as more people rate it.

The board moves on its own, too. GPT Image 2.5 Sunburst (max) sat at 1,199 when the Preview launched and reads 1,202 today — same weights, more votes, three points of drift. Three points is small next to fourteen, but it sets the scale for how seriously to take any single-figure difference in this family at one decimal place.

Screenshot of the Artificial Analysis AA-Image-Editing v2.1 leaderboard captured on 20 November 2026: MAI-Image-2.7 first at 1,215 Elo from 5,478 samples and $34.82 per 1,000 images, GPT Image 2.6 Sunburst (max) second, and MAI-Image-2.7-Flash seventh at 1,179 Elo from 5,891 samples and $10.24 per 1,000 images, with no MAI Image 2.7 Preview row anywhere on the board.

What the Preview could not do

The October release was deliberately partial, and Microsoft said as much at the time: editing first, generation later. Flash ships with both modes, and the closures matter more than the Elo step:

• Editing score — 1,193 Elo, No. 2, 4,891 votes in October vs 1,179 Elo, No. 7, 5,891 votes in November

• Generation — none: the Preview could not generate at all vs 1,158 Elo at No. 11 on AA-Image-T2I v2.1

• Price — unpublished; the October post carried no rate card vs about $10 per 1,000 images, vendor-reported, with Artificial Analysis independently estimating $10.24

• Access — MAI Playground and two leaderboards vs public preview on Microsoft Foundry, MAI Labs, and several third-party platforms

• Documented limits — nothing published vs the family model card's 32K-token context and 4,194,304-pixel output ceiling, about 2048×2048

• Transparency — unstated vs transparent backgrounds, confirmed for both 2.7 models in Microsoft's launch post

One caveat in that fifth line: the model card is titled for both models, but its spec rows describe MAI Image 2.7 and never break the limits out per variant. Publishing a limit "for the family" is not the same as publishing it for Flash, and Microsoft has not done the second thing.

Fourteen points against a nineteen-point interval

Current rows on the editing board carry 95% intervals of roughly ±12 to ±14. Flash's own band runs 1,166 to 1,192. The Preview was scored on a smaller vote pool, so its band was no tighter; comparing two independent estimates needs a wider interval than either alone — around ±18 if the October row carried a comparable band — and 14 points does not clear that.

Which means the honest reading of "Flash scores below the October Preview" is: the two are statistically indistinguishable in editing, on different dates, with different vote counts. Anyone who wants to argue Flash regressed has to explain why a model that also gained generation, a price and a deployment path would have lost editing quality in three weeks, and the numbers do not support the claim strongly enough to make it.

Two-column scoreboard comparing MAI Image 2.7 Preview with MAI Image 2.7 Flash across six dimensions: released 27 October 2026 versus 17 November 2026, editing Elo 1,193 at No. 2 versus 1,179 at No. 7, editing votes 4,891 versus 5,891, generation none versus 1,158 at No. 11, price unpublished versus about $10 per 1,000 images, access Playground and leaderboards only versus Foundry public preview and MAI Labs. A footer credits Elo and votes to Artificial Analysis and price to Microsoft, vendor-reported.

The number that actually separates them

Set the Elo aside and the split is stark: $10 per thousand images against nothing published at all. Flash's rate is Microsoft's own approximation — "~$10 per 1,000 images" in the launch post — and it is the first time anyone outside Redmond could price a 2.7-generation image. Artificial Analysis's derived figure of $10.24 per thousand lands close enough to corroborate the order of magnitude, which is more than can be said for most vendor rate claims in this category.

An approximation is still not a rate card. It answers "what does one default-size image cost" and not "what does a request carrying a stack of reference images and a 2048×2048 output cost," and Microsoft publishes no per-token or per-megapixel breakdown that would let a pipeline budget the second question. For a model pitched at production workloads, that is the gap to watch rather than the Elo.

Screenshot of Microsoft's MAI Image 2.7 launch post dated 17 November 2026, headlined 'Frontier quality at a fraction of the cost', showing the sentence that MAI Image 2.7 costs about $35 per 1,000 images while Flash costs about $10 per 1,000 images in public preview on Microsoft Foundry, above a quality-price Pareto chart plotting both models against GPT Image 2.6 Sunburst, MAI Image 2.6 and Krea 3 Ultra.

Neither MAI model is in the OrcaRouter catalogue as of 20 November 2026, and this piece will not pretend otherwise — the catalogue check comes back empty for the whole MAI image family. What is routable today is OpenAI's previous image generation: GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare, at the provider's list rate with 0% markup, which is a 1,202-Elo editor on the same board where Flash reads 1,179. When a MAI image model does reach a provider we route, the pass-through rule means the vendor's number becomes the number on our side the same day — a Microsoft price cut would arrive as a price cut, not as a renegotiation.

Canvas is the part nobody benchmarks

Flash has one job the Preview never had and the flagship does not have either: it powers Canvas, the interactive workspace Microsoft launched in MAI Labs the same day. Prompt or sketch in, diagram out, then click any part of the image to replace, move or restyle it in place. That is a product decision, not an evaluation result — Microsoft put the cheap, fast model behind the interactive surface and the expensive one behind the quality claims, which tells you what the company thinks each tier is for.

It also tells you where Flash's numbers came from. A model tuned for interactive editing inside a canvas is optimised for latency and iteration count, and neither of those appears on a leaderboard that scores a single edit per vote. Whatever Microsoft traded away to get there — steps, guidance scale, a smaller refinement pass — is not documented, and the model card's only description of Flash is a single clause: "an optimized version of MAI Image 2.7."

If you spent October in the Playground

The practical instruction is short. Re-run the prompts. The interface transferred, the editing task list transferred, and the price finally exists — but the checkpoint behind your old outputs is not on the board any more, and a prototype that taught you "temperature above 0.7 breaks text rendering" taught you that about a model you can no longer reach.

What did carry over is the shape of the work: object removal, replacement, attribute changes, inpainting, text substitution, artifact cleanup, enhancement. Those are the same seven editing actions Artificial Analysis scores, and they are the same ones the October post listed. A workflow built around them still holds; a tuning table built around the Preview's quirks does not.

Two signals would close the rest of it. The first is whether the Preview row stays retired — if Artificial Analysis keeps it off the board, the evaluator has effectively declared the two incomparable and every "Preview versus Flash" number becomes folklore. The second is a rate card with line items, which is what turns $10 into something a finance team can sign. Until then, the defensible statement is the narrow one: Flash is the shipping model, it edits as well as the prototype within measurement error, it generates too, and it costs about a cent an image.