MAI Image 2.7 Preview vs Meta Muse Image: Two Bets That Editing Is a Reasoning Problem

Both companies arrived at the same thesis from opposite directions. Meta Muse Image, which Meta put into its apps on July 7, 2026 and behind an API in late August, is built around the idea that a good edit requires a plan: it reasons about the request, can search the web for a missing fact, can run code to check a measurement, and then renders. MAI Image 2.7 Preview, released by Microsoft AI on October 27, 2026, makes the same argument with different evidence — it claims the top rank in Artificial Analysis's reasoning-based editing category, and it publishes no reference-image limit, no resolution, no rate card and no endpoint at all.

So the interesting question is not which one scores higher. It is what each company is willing to show you about the reasoning it is selling. Meta ships the machinery as visible tools and charges a penny an image. Microsoft ships a leaderboard placement and asks you to take the rest on trust. Those are different products even before you get to the fact that only one of them can be called from code.

The Same Claim, Made Two Ways

Meta's framing is agentic. Muse Image does not treat an edit as a single diffusion pass with a longer prompt; it treats it as a task with steps. Ask it to move a product onto a different surface and it may plan the geometry, look up what that surface actually looks like under a given light, and composite. That is why the launch material talks about search grounding and code execution as features rather than as safety infrastructure. They are the reasoning, made external.

Microsoft's framing is statistical. MAI Image 2.7 Preview's headline is a category sweep on a blind-voting board: reasoning-based edits moved from seventh place to first between 2.6 and 2.7, scene and style edits also from seventh to first, text and symbol edits from sixth to first. Microsoft says the model reaches the frontier in enhancement and restoration, reasoning-based edits, scene and style, and text or symbol editing, and that its largest single lead is 24 Elo in scene and style over the next-best model. Those are vendor statements about a third-party board, which is a more honest form of claim than a home-made benchmark — but it is still a claim about aggregate human preference, not about the steps the model took.

The difference matters when an edit fails. With Muse Image you can read the plan, see that it searched for the wrong product variant, and correct the intermediate step. With MAI Image 2.7 Preview you get a better or worse output and no account of why. For a one-off creative job that is fine. For a pipeline where failures have to be diagnosed, a visible plan is worth more than 40 Elo.

What Is Actually Documented

The spec lines are unusually lopsided, and the lopsidedness runs in Microsoft's disfavour on almost every axis a buyer plans around.

• Release — Meta Muse Image was announced July 7, 2026 and reached the Meta Model API in late August 2026; MAI Image 2.7 Preview was announced October 27, 2026.

• Scope — Muse Image generates and edits; MAI Image 2.7 Preview edits only, with text-to-image listed as coming soon.

• Access — Muse Image is in the Meta AI app, meta.ai, Instagram Stories in the US and WhatsApp, plus the vendor's own API and several third-party platforms; 2.7 is in MAI Playground, Arena and Artificial Analysis, with no API and no Foundry deployment.

• Price — Muse Image is a flat $0.01 per generated image, with filtered and failed images not billed; Microsoft has published no rate card for 2.7 in any currency.

• Reference images — up to ten for Muse Image; unpublished for 2.7.

• Output size — Muse Image tops out around 1,600 pixels; unpublished for 2.7, and the predecessor MAI-Image-2.6 capped at 2,359,296 pixels, roughly 1536×1536.

• Weights — neither is open.

• Independent score — Muse Image sat at 1,110 Elo on Artificial Analysis's image-editing board in an early-September 2026 snapshot; 2.7 debuted at 1,193 on AA-Image-Editing v2.1, evaluated 22 October 2026.

That last line is the one that gets quoted, and it is the one least safe to quote. The two scores come from different snapshots of a shifting board, so the 83-point gap between them is not a measurement — it is two measurements taken weeks apart on a leaderboard whose composition changed in between. Anyone telling you 2.7 beats Muse Image by 83 Elo is subtracting across runs.

The Penny and the Preview

Meta's price is the most aggressive number in consumer image generation right now: one cent per image, flat, no resolution tiers, no separate charge for the search grounding that makes the model interesting. At that rate a thousand edits cost ten dollars, which is less than the coffee budget of the meeting where you decide to run them.

Microsoft's price is not high; it is absent. There is no per-image figure for MAI Image 2.7 Preview because there is no billing surface to attach one to. The only honest cost model for it today is labour: every image is a manual round trip through Playground or a vote in Arena, so a two-hundred-image evaluation is a working day of someone's attention rather than a line item. For a design team testing whether the scene-and-style claim holds on its own assets, that trade is excellent — a day of attention for the second-best editing score in the world is cheap. For anyone building a product, the same trade is fatal, because the marginal cost of image two hundred and one is another manual round trip.

There is a second cost hiding in Meta's penny, and it is not financial. Meta's outputs carry Content Seal watermarking, and reporting in the months after launch found the detector failing to verify a majority of cropped watermarked images. If provenance matters to your workflow — and for advertising, licensing or journalism it does — then the watermark is neither a guarantee nor an obstacle, and you should decide what you need before you build on it rather than after.

Distribution Is the Part the Scoreboard Cannot See

The two models are being distributed into completely different worlds, and that predicts their futures better than any Elo.

Muse Image went straight into surfaces with billions of users: Instagram Stories, WhatsApp, the Meta AI app. That is an image model optimised for a person typing a sentence and getting something postable, where a one-cent price and a fast, agentic edit loop are exactly right. The API came later, and it came as an OpenAI-compatible endpoint — same request shape, different base URL — which tells you Meta expects developers to try it by changing two lines rather than by rewriting an integration.

MAI Image 2.7 Preview is in the opposite position. Microsoft has a developer platform, Foundry, and 2.7 is not on it. The model is being shown rather than sold, in the same three venues — Playground, Arena, Artificial Analysis — that its predecessors used. Two of those are evaluation surfaces and one is a demo. Whatever Microsoft's commercial plan is, it has not been published, and the gap between "second on the board" and "not in the catalogue" is the whole story of this launch.

What You Can Call Today

If the question is what to put behind an endpoint this week, the answer is neither of them at list-price parity. Meta Muse Image is callable through the vendor's own API and several third-party platforms at roughly a cent an image. MAI Image 2.7 Preview is not callable at all, and no amount of leaderboard position changes that.

What is callable, and sits immediately around 2.7 on the same board, is the pair of GPT Image 2.5 tiers — Sunburst at 1,199 and Flare at 1,182 on AA-Image-Editing v2.1 — and both are on OrcaRouter at their provider list prices with nothing added on top. That matters here for a specific reason: the two models this article is about bracket a real gap, and the only way to close it on your own assets is to run the same brief through models you can actually reach. One key over the image models we route, 0% markup on the provider price, and automatic failover if a provider degrades is a cheaper way to answer "does the reasoning claim survive contact with my product shots" than waiting for a rate card.

Which Bet to Take

If you are a designer or a creative lead, spend an afternoon in Arena this week comparing MAI Image 2.7 Preview against Muse Image on your own material. It costs nothing, the 2.7 restyling claim is the strongest single claim in the launch, and blind preference on your assets is the only evidence that transfers. Ignore the 83-point gap; run the comparison yourself.

If you are an engineer, Meta Muse Image is the one of the two with a price, a reference-image ceiling and an endpoint, and at a cent an image the experiment is nearly free. Watch the watermark question and watch the resolution ceiling — 1,600 pixels is a thumbnail and a hero image, not a print layout.

And if you are waiting for Microsoft: the signal to watch is not another leaderboard position. It is a Foundry deployment string or a published per-image rate for MAI Image 2.7 Preview. Until one of those exists, the model is a very good answer to a question nobody in production is allowed to ask.