MAI Image 2.7 Preview vs FLUX 3: A Leaderboard Placement Against a Price List

One of these models has a published price for every resolution it offers and can be called with an HTTP request. The other has a better score and no endpoint at all. MAI Image 2.7 Preview, released by Microsoft AI on October 27, 2026, sits second on Artificial Analysis's editing board at 1,193 Elo. FLUX 3, from Black Forest Labs, shipped its image tier on October 1, 2026, at $0.048 for a one-megapixel output and $0.607 for a sixteen-megapixel one, and charges the same whether you ask it to generate or to edit.

That asymmetry makes this less a head-to-head than a question about what you are buying. If the purchase is "the best editor I can evaluate this week," MAI Image 2.7 Preview wins on the strength of a score you cannot reproduce in your own pipeline. If the purchase is "an editor I can put behind a product on Monday," FLUX 3 is not competing with it — it is competing with everything else you could call instead, and the Microsoft model is simply not in the room.

Both facts are true at once, and which one matters depends entirely on where you are in a build.

What Each One Actually Is

The two models are not the same class of object, and the spec lines show it before any benchmark does.

• Scope — MAI Image 2.7 Preview is an editor only, with text-to-image listed as coming soon; FLUX 3 is a unified multimodal family spanning image, video with native audio, and robot action prediction, and its image tier does generation and editing through one endpoint with no mode switch.

• Release — MAI Image 2.7 Preview on October 27, 2026; the FLUX 3 family was announced July 23, 2026, with the image tier arriving October 1.

• Maximum output — MAI has published no resolution for 2.7; FLUX 3 Image scales to 4K, listed at 3840×2176 in Black Forest Labs' documentation.

• Reference images — unpublished for 2.7; up to ten for FLUX 3 Image.

• Spatial control — Microsoft describes scene restyling, relighting and text or symbol edits, with no published mask or box interface; FLUX 3 Image exposes bounding-box subject positioning and pixel-precise partial edits as documented parameters.

• Price — no rate card for 2.7 in any currency; FLUX 3 Image is $0.041 at 768 square, $0.048 at 1K, $0.100 at 2K and $0.607 at 4K, billed per image and identical for generation and editing.

• Weights — neither is open at full size, but Black Forest Labs ships FLUX 3 Image Klein, a 5.6-billion-parameter open-weights variant with the same request format, at $0.008 to $0.036 depending on resolution.

The Resolution Argument Is Really a Downstream Argument

FLUX 3's 4K ceiling sounds like a spec-sheet flex until you follow what happens to an edited image after the model returns it. A 2K edit that goes into a print layout, a large-format display, or a crop-heavy social pipeline gets upscaled somewhere downstream, and upscaling an edit is how you reintroduce the artifacts the edit just removed — softened text edges, a relight that loses its shadow gradient, a restored detail that turns to mush at 200%.

Sixteen megapixels at $0.607 is expensive per image and cheap per usable output when the alternative is a cheaper edit plus a separate upscale pass that has to be reviewed again. That arithmetic runs the other way for a marketplace thumbnail, where 768 square at four cents is the correct answer and 4K is waste.

Microsoft has not published 2.7's output resolution, so there is no honest way to place the two on this axis. The predecessor, MAI-Image-2.6, capped output at 2,359,296 pixels — about 1536×1536 — and it would be an inference, not a fact, to assume 2.7 matches it.

Ten References, or a Number Nobody Has Published

Reference-image count is the quiet determinant of whether an editing model fits a real workflow. Product photography needs a shot of the item, a shot of the surface it will sit on, and a style reference; brand work adds a logo lockup and a previous campaign frame. Five references covers the first case, ten covers both.

FLUX 3 Image takes up to ten. For MAI Image 2.7 Preview, Microsoft has disclosed nothing — not a maximum, not a mask convention, not a quality tier structure. That is not a criticism of the model's capability; it is a statement about what a buyer can plan around. An unpublished limit is a limit you discover during integration.

The same gap runs through spatial control. Bounding boxes are the least glamorous feature in image generation and the one that makes layout work reproducible: put the product here, put the headline there, keep the background. MAI's published description of 2.7 emphasises judgment-heavy edits — restyle the scene, fix the lighting, rewrite the sign — and says nothing about coordinate control. Those are different tools for different jobs, and only one of them is documented.

The Price of an Edit, Worked Through

Take a realistic batch: two hundred product shots a month, edited at 2K, with a tenth of them destined for print at 4K.

On FLUX 3 Image that is 180 images at $0.100 and 20 at $0.607 — $18.00 plus $12.14, or roughly $30 a month before any retry overhead. Retries are the number that actually moves the bill, because a model that needs three passes to get a label right costs three times the list price. Black Forest Labs returns the computed cost in the API response for each request, which makes that accounting straightforward rather than estimated.

On MAI Image 2.7 Preview the same batch costs nothing and takes a working day of someone's attention in Playground, because each image is a manual round trip. At small volume the labour is trivial and the free access is a genuine advantage — this is exactly the workload a preview is good for. At two hundred images a month the labour exceeds the entire FLUX 3 invoice.

This is where a gateway earns its place rather than being an extra hop. Both of the OpenAI editing tiers that bracket MAI Image 2.7 Preview on the leaderboard — GPT Image 2.5 Sunburst at 1,199 and GPT Image 2.5 Flare at 1,182 — are callable through OrcaRouter at their provider list price with nothing added on top, so the comparison that matters to a buyer is one key and one invoice across the models you are actually testing. The editorial line in this piece is that 2.7 is unavailable; the routing line is that the 42-point spread between the OpenAI tiers is a real choice you can measure on your own assets today.

Where FLUX 3's Own Numbers Come From

Black Forest Labs' launch benchmarks are vendor-reported and, importantly, mostly about video: a 77% preference win over Runway Gen-4.5 and 93% over Luma Ray 3.2 on ten-second 720p clips. Those are unreproduced head-to-head preference studies run by the company that made the model, and they say nothing about editing quality.

What FLUX 3 Image does not have is a public editing-board score comparable to 2.7's 1,193 Elo. So the scoreboard in this article is genuinely lopsided: one side has an independent rating and no product, the other has a product and no independent rating. Anyone who tells you one beats the other on editing quality is extrapolating from either a video benchmark or a resolution figure.

Klein Is the Part of FLUX 3 That Changes the Decision

The 5.6-billion-parameter Klein variant is open-weights, takes the same request format as the hosted model, and runs from $0.008 an image at 768 square. That combination does something MAI Image 2.7 Preview cannot do at any price: it lets a team run the model on its own hardware, keep assets inside its own perimeter, and iterate without a per-image invoice.

Open weights are also the reason FLUX 3 belongs in a different procurement conversation. A preview in someone else's playground is a capability demo. A downloadable checkpoint is an architecture decision, and it survives the vendor changing its pricing.

Which One to Pick

If you are evaluating editing quality and want to know whether the scene-and-style claim holds, use MAI Image 2.7 Preview in Arena this week. It is free, it is the second-best score on the board, and a blind comparison against a known model is the fastest way to find out whether the 24-point lead in restyling applies to your material.

If you need an editor in a pipeline, FLUX 3 Image is the only one of the two that answers the phone, and its published resolution ladder, ten-reference ceiling and per-image pricing let you size the bill before you write code.

If you need an editor in a pipeline today and want the strongest score you can actually reach, neither is the answer — the models directly above and below 2.7 on that board are. That is an unsatisfying conclusion to a comparison article, and it is the accurate one.