MAI Image 2.7 Preview vs HY Image 3.5: The 105-Point Gap You Should Not Compute

Put the two numbers side by side and the comparison looks settled. MAI Image 2.7 Preview scored 1,193 Elo on Artificial Analysis's image-editing board. HY Image 3.5 scored 1,088. That is a 105-point gap, which on an Elo scale is not a rounding difference — it is a different tier of model. The only problem is that the two figures come from different snapshots of a board that Artificial Analysis has been rebuilding, and subtracting across them is the same mistake as comparing a temperature in Celsius to one in Fahrenheit and calling the difference a heatwave.

That matters because the invalid subtraction is the most convenient argument available to either side. Microsoft benefits from a number that makes its preview look a hundred points ahead of a shipping competitor. Tencent benefits from a comparison against a model nobody can call. Neither has to make that argument for it to spread, because the figures live on the same leaderboard under the same name.

What follows is what each model actually offers, separated from the arithmetic that does not hold.

Two Snapshots Under One Board Name

MAI Image 2.7 Preview's 1,193 comes from AA-Image-Editing v2.1, snapshot dated October 22, 2026. It placed second, behind GPT Image 2.5 Sunburst at 1,199 and ahead of GPT Image 2.5 Flare at 1,182.

HY Image 3.5's 1,088 comes from a September 2026 reading of the editing board: fourteenth of ninety-seven models, with a confidence interval of 1,079 to 1,097 and 5,550 pairwise comparisons behind it. HunyuanImage 3.0 Instruct sat at 1,066 on the same reading and Qwen-Image-2.1 at 1,074.

The board version and the evaluation window differ, which is enough to make the difference between the two figures uninterpretable. The existing post on this blog about MAI-Image-2.6 already documents why: when the board was rebuilt as v2.1, the older model's score moved, and a "regression" that looked like a capability loss was an artifact of the ruler changing. The same caution applies in the other direction. A 105-point gap assembled from two rulers is not a measurement of anything.

What can be said with confidence is narrower and still useful: both models are in the top fifteen of the editing board at the time each was measured, and neither is at the top.

Where HY Image 3.5 Is Straightforwardly Better

Three things, and none of them are close.

• You can call it. HY Image 3.5 is available through Tencent Cloud's API, with a documented endpoint taking the Chat or Messages protocol and an image-generation guide that lists every parameter. MAI Image 2.7 Preview has no API, no Foundry listing and no published price.

• You can price it. Tencent charges ¥0.15 per 2K image, roughly two and a half cents, and bills only for generated output. Microsoft has published no rate card for 2.7 in any currency.

• It does the whole job. HY Image 3.5 handles text-to-image, image-to-image and multi-turn editing through one endpoint. MAI Image 2.7 Preview is editing only, with generation listed as coming soon.

That third point is the one that changes what kind of product each is. Multi-turn editing means the model holds a subject across a sequence of revisions in one conversation — the working pattern for a designer iterating on an asset over an afternoon. A single-pass editor that starts fresh on every request is a different tool, and Microsoft has not said which pattern 2.7 supports.

The Text-to-Image Board Turns the Ranking Upside Down

Here is the detail that should make anyone cautious about single-number comparisons: on Artificial Analysis's text-to-image board, HY Image 3.5 scored 975 Elo and placed sixty-sixth of 166 models. That is roughly twenty points below HunyuanImage 3.0 Instruct, its own predecessor, and it is the opposite of the picture the editing board paints, where 3.5 sits twenty-two points above the same predecessor.

A model that beats its predecessor on one board and loses to it on another is not a contradiction in the model. It is a signal that the two evaluations are measuring different populations of prompts, and that a model tuned for instruction-following edits can regress on open-ended generation without any of its numbers being wrong. Tencent's own framing supports this reading: the launch material is about professional creative work, text rendering, editing consistency, and a chain-of-thought prompt rewriter, with the text-to-image mode presented as one capability among several rather than the headline.

For MAI Image 2.7 Preview the corresponding board is empty, because the model cannot generate. That is not a missing data point; it is the product decision.

Different Failure Modes, Not Different Quality Levels

The categories where each model claims strength point at genuinely different problems.

Microsoft's claim for 2.7 is whole-frame judgment: restyle the scene, relight it, restore it, rewrite the text so it sits inside the image rather than on top of it. Artificial Analysis has the model first on four of seven editing actions, with the widest lead in scene and style at 24 points, and Microsoft says it improved across ten use cases and seven actions relative to 2.6.

Tencent's claim for HY Image 3.5 is consistency across revisions and legibility of rendered text, including small and multilingual text, plus realism. Its evidence is an internal blind test: about a 30% higher win rate than Hy Image 3.0 across hundreds of professional designers, roughly level with Seedream 5.0 Pro, and slightly ahead of Nano-Banana Pro and Qwen-Image-3.0 Pro. Those are vendor-run GSB results, not independent rankings, and should be read as such.

Put together, the two claims describe a real division of labour. If the task is "make this photograph look like a winter campaign," 2.7's scene-and-style margin is aimed at exactly that. If the task is "keep this product identical across forty renders and keep the label readable," HY Image 3.5's stated strengths are aimed at that. Neither claim is evidence about the other task.

What Each Costs at Volume

The pricing asymmetry is large enough to decide most real deployments on its own.

HY Image 3.5 at ¥0.15 per 2K image is about $0.024. A thousand images a month is roughly $24, before retries. Output resolution is documented at 2K, with the API's size parameter accepting a much wider range — width and height from 256 to 8192 under a 16,777,216-pixel area cap, which covers 4096×4096 at the top. The reference ceiling is five per request in Tencent's launch announcement; the API guide's parameter table documents zero to three for the older 3.0 endpoint and does not restate a number for 3.5, so resolve that against the live documentation before you plan a workflow around it.

MAI Image 2.7 Preview costs nothing and consumes the most expensive resource in the building: human attention in a browser tab. A thousand images is a thousand manual round trips. The model is free in the sense that a sample is free.

If a team wants one of these in a pipeline this quarter, the decision is already made by availability, and it is worth stating plainly that OrcaRouter does not host either model — HY Image 3.5 is reachable through Tencent's own cloud API, and MAI Image 2.7 Preview is reachable through nothing. Where a gateway does change the calculus is in the fallback layer: if you build on a preview-grade editing endpoint, a routing config that fails over to a second model on error or timeout is the difference between a bad afternoon and a broken release.

A Test That Would Settle It

Leaderboards are built from neutral prompts and single passes. The question a buyer actually has is what happens on the fifth revision of the same asset, with a reference image and a constraint that must survive every step.

HY Image 3.5 can be tested that way today: send a source image, a reference, and a multi-turn sequence, and check whether the subject drifts by turn four. MAI Image 2.7 Preview can be tested in Playground for single-pass quality and in Arena for blind preference, but not for sequence consistency, because there is no sequence interface.

Until Microsoft ships an endpoint and Artificial Analysis re-reads both models on the same board version, the honest comparison is this: Tencent has the cheaper, callable, multi-turn model with a mid-table score, and Microsoft has the better-scoring single-pass editor that nobody can deploy. The 105-point gap between them is a coincidence of timing, not a finding.