Images 2.6 vs MAI-Image 2.7 Preview: One Is Ranked Second, the Other Is Callable

There is a version of this comparison that a benchmark table makes look simple: MAI-Image-2.7 Preview scores 1,196 Elo on Artificial Analysis's AA-Image-Editing v2.1 board, good for second place, while Images 2.6 — OpenAI's rollout of GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare on 4 November 2026 — has no row on that board at all. Ranked model versus unranked model. The problem is that the ranked model cannot be called. MAI-Image-2.7 Preview has no public API, no Azure AI Foundry listing, no published price and no open weights. It exists in Microsoft's MAI Playground, on Arena, and as a name in a leaderboard column.

So the useful question is not "which is better." It is "which of these can you put in a product this week, and what are you giving up by picking the one that answers the phone." That answer is lopsided, but not in the direction the Elo numbers suggest.

Three things that are true on 5 November 2026

• MAI-Image-2.7 Preview is an editor and only an editor. It requires a source image. Text-to-image is described by Microsoft as "coming soon," with no date attached — which is why you will find no MAI-Image-2.7 Preview row on the AA-Image-T2I board and why the 1,196 figure is an editing score that says nothing about generation.

• Images 2.6 does both, and adds a third output type. Generation, editing, and — new on 4 November — single-step animated GIF creation inside ChatGPT from a prompt or an uploaded photo, plus SVG export and markup-guided edits in the API. It is a broader product with, as of today, no independent score of any kind.

• The two-point gap at the top is not really two points. MAI-Image-2.7 Preview's 1,196 (±15) sits three Elo below GPT Image 2.5 Sunburst's 1,199 (±11). Both confidence intervals are wider than the difference. Statistically these are level, and OpenAI's 2.6 successor is the model Microsoft would actually be measured against if anyone had benchmarked it.

What the 1,196 is actually measuring

Editing Elo is an average over a distribution of edit requests, and averages hide the thing buyers care about. Microsoft's own breakdown is more informative than the headline number: it says the model leads on four of seven editing actions — enhancement and restoration, reasoning-based edits, scene and style, and text or symbol editing — with its widest margin in scene and style at 24 points over the next-best model.

It loses the other three. Composition, individual-object edits and identity preservation stay with OpenAI. Read as a description of behaviour rather than a scoreboard, that is a clean split: give MAI-Image-2.7 Preview a frame and tell it to make the whole thing feel like dusk, or restore a damaged scan, or rewrite the shopfront lettering so it sits in the scene, and it is the strongest thing on the board. Ask it to move one object two feet left without altering the person holding it, and OpenAI's models have historically done that better.

The gain over its own predecessor is also worth noting because it is large by Elo standards: +45 over MAI-Image-2.6, which sits at 1,151 with nearly 17,000 samples. That is not noise. Microsoft shipped a real improvement on 27 October, nine days before OpenAI shipped 2.6.

The pricing asymmetry is the whole decision

You cannot buy MAI-Image-2.7 Preview. There is no rate card, and Microsoft has not said when there will be one. The nearest proxy is what the previous generation costs on Foundry: MAI-Image-2.6 bills $5 per million text input tokens, $8 per million image input tokens and $38 per million image output tokens, which works out to roughly 3.9 cents for a typical image; Artificial Analysis lists it at $38.9 per thousand images and the Flash variant at $19.5.

Images 2.6, by contrast, has a live price today. It is token-metered and unchanged from the 2.5 generation: $5 per million text input (cached $1.25), $8 per million image input (cached $2), $30 per million image output. Because OpenAI publishes no flat per-image figure, the practical estimate has to come from the output-token budget — about $0.053 at high quality, $0.094 at xhigh and $0.21 at max, with batch rates roughly half. Those budgets are inherited from 2.5, so treat the per-image numbers as an approximation and the per-token numbers as the contract.

Notice what that does to the comparison. The cheaper model on paper — MAI-Image-2.7 Preview, presumably priced near its predecessor's four cents — is infinitely expensive right now, because the only way to use it is to open a playground and work by hand. The model with the higher per-image ceiling is the one with an invoice you can actually receive.

This is the case where a routing layer is not a cost optimisation but a way to keep the decision reversible. OrcaRouter carries OpenAI's image line at provider list price with 0% markup — GPT-Image-2.6 Sunburst and Flare sit beside GPT-Image-2.5 Sunburst and Flare, GPT Image 2, Imagen 4 and Gemini's flash-image models on one key, with no second contract and no second SDK. When Microsoft does publish a 2.7 endpoint, adding it is a catalogue change on our side and a model-string change on yours, not a migration. Failover, meanwhile, is what stops a nine-day-old preview from being a production dependency the day it rate-limits.

What you can and cannot conclude today

You cannot conclude that MAI-Image-2.7 Preview beats Images 2.6 at editing, because nobody has scored 2.6. You also cannot conclude the reverse. The only defensible reading of the boards is that Microsoft has closed to within statistical noise of OpenAI's previous generation on editing, and that OpenAI's current generation is unmeasured.

You can conclude three narrower things, and they are enough to act on.

First, if your problem is whole-scene transformation — relighting, restyling, restoration, making text sit inside a photograph rather than on top of it — MAI-Image-2.7 Preview is, by Microsoft's own breakdown, the better-shaped tool, and it is the one place a playground trial is worth the afternoon. There is no charge to evaluate it, which is unusual and useful.

Second, if your problem is a literal instruction applied to a specific region — move this, remove that, keep the face identical — the seven-action split says OpenAI's line is where that work belongs, and GPT-Image-2.6 Sunburst is the current build of it.

Third, if you need generation, animation, SVG or an API, the comparison collapses. MAI-Image-2.7 Preview offers none of the four.

What would change this

Two events, either of which would make this article obsolete within a month.

A Foundry listing with a price. Microsoft has shipped 2.6 and 2.6-Flash into Foundry already, so the pattern is established and the delay on 2.7 looks like a preview-to-GA gap rather than a policy. If a 2.7 endpoint lands at roughly the 2.6 rate card, it becomes the default recommendation for scene-level editing work overnight — a top-two editing score at four cents an image, with no reference-image surcharges, undercuts Krea 3's $0.06 base and its $0.07-per-reference add-on by a wide margin on multi-reference jobs.

An AA row for Images 2.6. The image boards add models in batches rather than day by day — several joined in the thirty days before this was written — so the lag is normally weeks rather than months. When 2.6 appears, the number that matters is whether it clears 1,199 — its own predecessor — because a successor that regresses on editing while adding GIFs would be a feature release, not an editing release, and the 1,196 that MAI-Image-2.7 Preview posts would stop being the second-best score and start being the best available one.

Until one of those happens, the honest summary is short. The best editing model you can benchmark is not the best editing model you can buy, and the model you can buy has not been benchmarked. Anyone telling you which of these two wins is guessing.

The practical sequence

If you are evaluating this pair for real work, the sequence that wastes the least time is: run your ten hardest whole-scene edits through MAI-Image-2.7 Preview in the playground, because it costs nothing and it is genuinely good at that shape of task; then run the same ten through GPT-Image-2.6 Sunburst on the API, where the cost of the test is roughly fifty cents; then run them through GPT-Image-2.5 Sunburst as the control, because that is the model with 18,300 samples behind its 1,199 and the only one of the three whose position you can trust. If 2.6 beats its predecessor on your ten, it is probably a real improvement. If it does not, you have learned something the leaderboards cannot tell you yet, for under two dollars.

The one thing not to do is defer the decision until the benchmarks arrive. Microsoft's preview has no announced endpoint and OpenAI's successor has no announced score, and both of those gaps could close next week or next quarter. The models are available enough to test now, which is more than can be said for the evidence.