MAI-Image-2.7-Preview vs GPT-Image-2.5: The Sharper Editor vs the One You Can Call

Start with the job, not the leaderboard. If your product edits a catalogue of product shots overnight through an API, the choice between MAI-Image-2.7-Preview and GPT-Image-2.5 is not close, because only one of them accepts a request. If you are judging the edit itself — relighting a scene, replacing a label, keeping a character consistent across five rounds — Microsoft's new model has a real claim to the top, and it makes that claim with a score line thin enough to argue about. The matchup is therefore not "which model is better." It is "how much editing quality is a production API worth."

Microsoft released MAI-Image-2.7-Preview on October 27, 2026. It is an editing-only preview, reachable through MAI Playground, Arena and Artificial Analysis, with no Microsoft Foundry deployment and no published price. OpenAI's GPT-Image-2.5 has been generally available since September 8, 2026 in two variants — Sunburst for quality, Flare for speed — with a public rate card, text-to-image and editing, and a documented API. That asymmetry frames every number below.

The numbers, and how much weight they carry

On Artificial Analysis's AA-Image-Editing v2.1 board, evaluated October 22, 2026, GPT-Image-2.5 Sunburst (max) leads with 1,199 Elo. MAI-Image-2.7-Preview is second at 1,193. GPT-Image-2.5 Flare (max) is third at 1,182.

• MAI-Image-2.7-Preview vs Sunburst: 6 Elo behind, at the max quality tier

• MAI-Image-2.7-Preview vs Flare: 11 Elo ahead, at the max quality tier

• MAI-Image-2.7-Preview vs MAI-Image-2.6: 42 Elo ahead

• Arena Edit: Microsoft reports 2.7 at No. 3, and itself as the No. 2 lab

Those figures come from Microsoft's launch post, which reproduces the Artificial Analysis board; Artificial Analysis is the independent evaluator, Microsoft is the channel through which we are reading it. Elo gaps of six and eleven points are within the range where vote composition and confidence intervals can flip an ordering, so the defensible claims are narrower than the table suggests: 2.7 is competitive with the best GPT-Image-2.5 configuration and clearly ahead of the fast one, not that it is the single best editor in the world. The Arena No. 3 pointing the same direction is corroboration, not confirmation, because the two boards use different voters and scales.

One more caveat that matters for cost: the OpenAI numbers are the "max" tier. That is the most expensive setting in the family, and it is not the tier most production traffic runs at.

What each model actually is

• Modes: MAI-Image-2.7-Preview is editing only at launch, with text-to-image "coming soon"; GPT-Image-2.5 does text-to-image and editing through the same family.

• Variants: Microsoft ships one 2.7 preview; OpenAI ships Sunburst (quality-first) and Flare (latency-first) at identical token rates.

• Access: 2.7 is in MAI Playground, Arena and Artificial Analysis; GPT-Image-2.5 is a public API.

• Price: 2.7 has none published; GPT-Image-2.5 bills $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens, with cached input at $1.25 and $2.00.

• Quality ladder: OpenAI exposes low, medium, high, xhigh and max; Microsoft has not published an equivalent control for 2.7.

• Reference images: OpenAI supports multi-reference edits, up to 16 images by its own documentation; Microsoft published five for MAI-Image-2.6 and has not published a limit for 2.7.

The reference-image and resolution details are where the spec sheets get slippery. MAI-Image-2.6 capped output at 2,359,296 pixels — 1536×1536 — and accepted five reference images, both documented in Microsoft Foundry's API reference. Microsoft has not published 2.7's equivalents, and it is not safe to assume the 2.6 limits carried over unchanged. OpenAI's 2.5 family advertises higher resolution ceilings in its launch materials, though the published token examples top out at 1536×1024, so the practical ceiling depends on the size you request. If a fixed 4K deliverable is a hard requirement, verify it against the live API rather than either vendor's marketing page.

The cost of choosing the better editor

Run the arithmetic on 10,000 square images at 1024×1024. At max quality, GPT-Image-2.5 costs about $0.211 per image, or roughly $2,110. At high, about $0.053 each, or $527. At medium, about $0.013 each, or $132. Those are output-token estimates from OpenAI's published rates, not a quote, and edits add image-input cost on top because the reference image is billed as input.

MAI-Image-2.7 has no price, so the only anchor is its predecessor. MAI-Image-2.6 in Foundry costs $5 per million text input tokens, $8 per million image input tokens and $38 per million image output tokens, which Artificial Analysis estimated at roughly $0.039 per image — about $390 for the same 10,000. If 2.7 launches at a similar rate, it would sit between GPT-Image-2.5's high and medium tiers on cost while scoring above the Flare model. That is an "if," not a finding, and Microsoft has given no signal about where it will land.

There is a second cost trap in the OpenAI ladder. The tiers were relabelled in 2.5: the new high setting uses roughly the token budget of the old medium, and max uses roughly the old high. Teams that moved from gpt-image-2 with an unchanged quality parameter got a cheaper render at what used to be a lower quality setting. Check which tier your code actually requests before comparing a bill to last quarter's.

Where 2.7 is genuinely better

Microsoft's category claims are the useful part of its post. It says 2.7 improves on 2.6 across all 10 use cases and all 7 editing actions Artificial Analysis evaluates, reaches the frontier in enhancement and restoration, reasoning-based edits, scene and style, and text or symbol editing, and leads the next-best model by 24 Elo in scene and style. The rank jumps are large: reasoning-based edits and scene and style both moved from No. 7 to No. 1, and text or symbol edits from No. 6 to No. 1.

Those are the edits that most often fail in production. Relighting a product into a new environment, replacing signage without warping the perspective, restoring a low-quality asset without inventing detail — each requires the model to understand how the scene works, not just where to paint. A 24-Elo lead in scene and style, if it holds, is a bigger practical difference than the six-point overall gap suggests, because it is concentrated exactly where generic editors tend to produce obvious artifacts.

OpenAI's counter-argument is not a category score. It is multi-turn consistency: Sunburst is positioned for edit-heavy workflows where the same subject has to survive several rounds of changes, and OpenAI's own documentation recommends compositing approved edits back into the original when pixels must stay identical. That is an admission that even the best editor drifts. Microsoft has not published an equivalent multi-turn evaluation, so the one workflow that matters most for iterative design is the one with the least comparable evidence.

Running both without two contracts

This is where the matchup stops being symmetrical. GPT-Image-2.5 Sunburst and Flare are both on OrcaRouter, behind one API key and one billing relationship, at 0% markup — provider list price passed through, with automatic failover if a provider degrades. You can send draft iterations to Flare and final renders to Sunburst without opening a second account, and a routing rule can make that switch by task type rather than by code branch.

MAI-Image-2.7-Preview is not on OrcaRouter, and cannot be until Microsoft exposes a public API. Today it is a playground model: manual, human-in-the-loop, fine for evaluation and asset work, not something you can put behind a queue. That is not a knock on the model's quality; it is the difference between a benchmark winner and a dependency.

The call

• You need an edit endpoint this quarter: GPT-Image-2.5. Sunburst for final assets, Flare for iteration volume. The six-Elo deficit to 2.7 is smaller than the cost of not shipping.

• You need the single best edit on a handful of high-value assets: test MAI-Image-2.7-Preview in the playground. Scene restyling and text replacement are where Microsoft claims the widest lead, and those are worth a manual pass on hero imagery.

• You need text-to-image as well as editing: GPT-Image-2.5, full stop. 2.7 does not generate yet.

• You are cost-sensitive at volume: price Flare against your real edit sizes, not the headline per-image numbers. Image-input tokens dominate when you send several references.

The honest summary is that Microsoft has built the better editor and OpenAI still sells the better product. Those can both be true for as long as 2.7 stays a preview — and the day Microsoft puts a Foundry version and a price on it, this comparison gets a second act.