MAI-Image-2.7-Preview Lands Six Elo From the Top — and Still Without an API

Six Elo is all that separates Microsoft's newest image editor from the best one OpenAI sells. On October 27, 2026, Microsoft released MAI-Image-2.7-Preview, which debuts at No. 2 on the AA-Image-Editing v2.1 leaderboard with 1,193 points against 1,199 for GPT Image 2.5 Sunburst (max) — ahead of GPT Image 2.5 Flare (max) at 1,182, and 42 points clear of Microsoft's own MAI-Image-2.6. The score is not the strange part. The strange part is everything Microsoft did not ship with it: no Microsoft Foundry deployment, no published price, and no text-to-image mode.

That combination makes "Preview" the most important word in the model name. Microsoft has a model that independent human voting places within noise of the category leader, and the only ways to touch it are MAI Playground, the Arena Edit board, and Artificial Analysis. For a company that put MAI-Image-2.6 into Foundry with a rate card six weeks earlier, shipping the successor as a playground demo is a deliberate sequencing choice — and it changes what the ranking actually means for anyone building something.

What the board says, precisely

The numbers below are from Microsoft's launch post, which reproduces the Artificial Analysis AA-Image-Editing v2.1 board; the evaluation snapshot is dated October 22, 2026. Artificial Analysis is an independent evaluator, but the figures here reach us through Microsoft's chart, and we have not re-run the board ourselves. Treat the ordering as the signal and the exact gaps as provisional.

• GPT Image 2.5 Sunburst (max) — 1,199 Elo, No. 1

• MAI-Image-2.7-Preview — 1,193 Elo, No. 2

• GPT Image 2.5 Flare (max) — 1,182 Elo, No. 3

• MAI-Image-2.6 — 1,151 Elo, tied with Krea 3 (Ultra) and Nano Banana 2.1

• MAI-Image-2.6-Flash — 1,144 Elo

• GPT Image 2 (high) — 1,136 Elo

Two things about that list are easy to misread. First, 2.7 is not merely Microsoft's best editing model; it is the only non-OpenAI model within eleven points of the top, and it sits between OpenAI's two 2.5 variants. Second, MAI-Image-2.6 looks like it regressed. It did not. Artificial Analysis rebuilt the editing board as v2.1, and scores from the older board — where Microsoft's September post put 2.6 at No. 1 — cannot be subtracted from these. Within v2.1, 2.6 is a 1,151 model and 2.7 is a 1,193 model. That is the only comparison the numbers support.

On Arena, the other independent board, Microsoft reports 2.7 at No. 3 on Arena Edit, which it says makes Microsoft AI the No. 2 lab on that board. Arena and Artificial Analysis use different vote pools and different scales, so the No. 2 and No. 3 are not a contradiction — and neither should be converted into the other.

Where the 42 points came from

Microsoft's claim is specific: MAI-Image-2.7 improves on MAI-Image-2.6 across all 10 use cases and all 7 editing actions that Artificial Analysis evaluates. It says the model reaches the category frontier in enhancement and restoration, reasoning-based edits, scene and style, and text or symbol editing, with its widest margin in scene and style — 24 Elo ahead of the next-best model there.

The category-rank movement is the more concrete version of the same claim. Reasoning-based edits went from No. 7 to No. 1, scene and style from No. 7 to No. 1, and text or symbol edits from No. 6 to No. 1. Those are large jumps for a single version step, and they line up with the use cases Microsoft names in its own examples: restyling or relighting a whole scene, changing text and symbols while keeping them integrated, and restoring existing assets.

The caveat is structural. Category scores are cuts of the same preference data that produces the overall Elo, not seven additional independent experiments. If 2.7 wins overall, it will usually win most of the cuts. The informative part is the shape — a model that was mid-table on reasoning and text edits now leading them — not the individual rank deltas. And every capability statement in this section is Microsoft's characterization of an Artificial Analysis evaluation, which is a stronger form of evidence than a vendor benchmark but still not a controlled third-party reproduction of Microsoft's workflow claims.

A launch with the commercial half missing

MAI-Image-2.7-Preview is available today in MAI Playground, on Arena, and on Artificial Analysis. Microsoft's post says the company is "starting with editing, with text-to-image generation coming soon." That is the entire availability story. There is no MAI-Image-2.7 entry in Microsoft Foundry's model catalogue, no deployment version string, and no per-token rate published anywhere Microsoft controls.

The contrast with the predecessor is sharp. MAI-Image-2.6 entered Foundry on September 4, 2026, with MAI-Image-2.6-Flash alongside it, and a published rate card: $5 per million text input tokens, $8 per million image input tokens, and $38 per million image output tokens for the main model; $1.75, $2.50 and $19 for Flash. Artificial Analysis's per-image estimate for 2.6 worked out to roughly four cents, and about two cents for Flash. Those are not 2.7's prices — they are the only public anchor for what a Microsoft image edit has cost until now.

So the honest description of this launch is: best-ever Microsoft editing quality, available free in a playground, with no production contract. That is useful for evaluation and useless for a pipeline. Anyone who needs an edit endpoint this week is still choosing between MAI-Image-2.6 in Foundry preview and a competitor.

Why ship the editor first?

Microsoft has not explained the sequencing, so this is inference, not reporting. But the pattern is visible in the product surfaces. Editing is where Microsoft's image work actually lands: PowerPoint, OneDrive, Bing Image Creator and Copilot all need "change this, keep the rest" behavior far more than they need another text-to-image box. Editing is also the harder capability to differentiate and the easier one to demo against a specific brief — relight the scene, swap the label, restore the scan.

Text-to-image, by contrast, is crowded and expensive to serve. Holding it back lets Microsoft ship the capability with the clearest commercial hook while it finishes the generation side and, presumably, the Foundry packaging. The risk is that a preview with no price and no API does not convert into developer mindshare, however good the Elo is. The September 2.6 release showed Microsoft knows how to do the commercial half; the question is how long 2.7 stays half-shipped.

What to do if you need an image model now

Try 2.7 in MAI Playground against your five hardest edits. The categories Microsoft claims — relighting, text replacement, restoration — are exactly the ones that break in production, and a free playground is the cheapest way to see whether the Elo gain survives your own assets. Do not build a roadmap around it until it has a Foundry version and a price.

For production editing, the practical options are the two models that actually accept a request. MAI-Image-2.6 remains in Microsoft Foundry preview at the rates above. On the OpenAI side, GPT-Image-2.5 Sunburst and Flare are both live APIs, and both are routable through OrcaRouter on one key at 0% markup — provider list price passed through, so any OpenAI price change is live here the same day rather than waiting on a reseller update. Sunburst is the quality tier that holds the top of this board; Flare is the cheaper, faster variant at the same token rates.

That is the awkward shape of the current market: the model with the second-best editing score is not purchasable, and the two models that are purchasable are the ones it is chasing. OrcaRouter does not host MAI-Image-2.7-Preview, and we are not going to pretend otherwise.

What would move 2.7 from headline to option

Three things, in order of importance. A Foundry listing with a version string and a rate card would make it callable; a price at or below MAI-Image-2.6's would make it interesting for volume; and text-to-image parity would make it a replacement rather than a specialist. None of those is a research problem — they are packaging decisions, and Microsoft made all three for 2.6 in a single September release.

Until then, the six-Elo gap is a statement about how fast Microsoft's image team is closing on OpenAI, not a purchasing decision. The editing crown is now contested. The API key is still OpenAI's to lose.