Title card reading 'Kling Image 4.0 vs MAI Image 2.7 Preview' with the subtitle 'One has a rate card, one has a tab'. A left card headed Kling Image 4.0 lists 17 November 2026, 2K/4K/8K output and 6.1 to 15.4 cents a frame; a right card headed MAI Image 2.7 Preview lists 27 October 2026, editing only and no published price, with a small versus divider between them.
Guides & Insights

Kling Image 4.0 vs MAI Image 2.7 Preview: One Has a Rate Card, One Has a Tab

Author

Gideon Frost

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Kling Image 4.0 shipped on 17 November 2026 with an invoice attached: 16 units for a 2K still, 24 at 4K, 40 at 8K, plus 0.8 units for every reference you attach. MAI Image 2.7 Preview arrived on 27 October 2026 as an editing-only model with no published price, no published resolution and no API — you reached it through MAI Playground, Arena and Artificial Analysis, one image at a time, by hand. On 18 November the Preview's row is gone from Artificial Analysis's AA-Image-Editing v2.1 board; the model it was previewing, MAI-Image-2.7, sits first on that board at 1,215.17, and Kling Image 4.0 sits sixth at 1,182.89. The Preview's own last recorded score was 1,193, which is higher than Kling's. So the only model in this pair that you can put inside a pipeline is also the one with the worse number, and that inversion is the whole comparison. Every image model we route, with its provider's own list price, is listed in the OrcaRouter model catalogue.

Both are real releases rather than leaks. Kling's release history carries a 2026-11-17 entry headlined "Kling Image 4.0 Is Here!", its account posted the availability note in the early hours of 18 November, and the developer price sheet was live before either. Microsoft put the Preview on the editing board with a model card and an explicit "starting with editing, with text-to-image generation coming soon". Three weeks later the November model closed that gap and took the top slot. What did not appear, at any point, is a way for a developer to call the October model.

A withdrawn row is not a rival

Leaderboard rows are not permanent fixtures. Models get superseded, votes stop being collected, and the entry disappears — which is what happened to the Preview between its 27 October debut and 18 November. Reading the board today gives you MAI-Image-2.7 at No. 1 on 1,215.17, GPT Image 2.6 Sunburst at 1,210.84 and 2.6 Flare at 1,205.42 behind it, then the two GPT Image 2.5 tiers at 1,201.50 and 1,184.65, and Kling Image 4.0 at 1,182.89 in sixth. There is no Preview row to place anywhere in that list.

This matters for a practical reason rather than a pedantic one. A score attached to a model you cannot call is a data point about a research direction, not a purchasing signal. If you quoted 1,193 in an internal deck, you were quoting the October checkpoint of a system whose shipping version now scores 1,215 — and whose shipping version is a different product with a different feature set. The useful comparison, if you want a number at all, is Kling's 1,182.89 against MAI-Image-2.7's 1,215.17: a 32-point gap on the editing board, in favour of Microsoft, on a model that does have a price of roughly $35 per thousand images.

Against the Preview specifically, the honest statement is narrower. The Preview led Kling by about ten Elo in editing while it existed, from a smaller vote population, on a model with no generation mode. Ten points is inside the confidence band those rows carry. It is not a lead you can build on, and it is not a lead you can test, because there was never an endpoint to send a request to.

Screenshot of the Artificial Analysis AA-Image-Editing v2.1 leaderboard in English, with MAI-Image-2.7 first at 1,215 Elo, GPT Image 2.6 Sunburst second at 1,211, GPT Image 2.6 Flare third at 1,205, the two GPT Image 2.5 tiers fourth and fifth, and Kling Image 4.0 sixth at 1,183, with no MAI Image 2.7 Preview row anywhere on the board.

What the Preview was, precisely

Microsoft documented the October model as an editing surface with three named strengths: restyling a scene, relighting it, and changing text or symbols inside it. The availability sentence was equally narrow — Arena, Artificial Analysis, MAI Playground — and the Playground link itself still pointed at the previous model, which tells you how provisional the release was. No Foundry deployment, no rate card, no reference maximum, no resolution.

What that leaves is a model you evaluate by eye. That is not worthless: an art director comparing ten of their own images in a browser can learn a great deal in an hour. But it produces no cost estimate, no throughput figure and no integration path, and the three things a team needs before committing a production job are exactly those three.

Kling's side of the ledger has the opposite profile. Its five published capabilities — true-to-life realism, native 8K, small-text rendering, coherent image sets, long-image generation and editing — are vendor claims with no independent verification beyond the Elo score, but every one of them arrives with a unit price attached, which makes them falsifiable in a different way: you can deploy one and check the invoice.

The unit ladder, and the reference bill hiding in it

Kling prices in units and then converts. The still ladder is 16 units at 2K, 24 at 4K and 40 at 8K. At the entry package rate of $0.00384 a unit that is about 6.1 cents, 9.2 cents and 15.4 cents an image; the rate falls to $0.00381 in the $999 package and $0.00373 in the $4,000 one, which moves those numbers by less than three percent. Artificial Analysis lists the model at 6.3 cents, which sits inside that arithmetic and is a reasonable sign that the published unit price is the real one.

The line that changes budgets is not on the ladder. Reference images and reference elements are billed now — 0.8 units each, one element counted as one image — and Kling's own documentation notes that this "differs from previous models, where reference images and reference elements were free". For a set built from four references, each delivered frame costs 16 units plus 3.2 units of references, so a twelve-shot campaign at 2K is 230.4 units rather than 192: about 88 cents instead of 74, at the entry rate. The surcharge scales with how reference-heavy the work is, not with how much output you get, which means the workflows that need consistency most are the ones that pay most for it.

• Single still — 2K 16 units, 4K 24 units, 8K 40 units

• Reference input — 0.8 units per reference image or reference element, newly billable

• Long image, webpage — 140 units at 1K, 152 units at 2K

• Long image, e-commerce detail page — image-to-image only, 60 / 120 / 200 units at 1K for short, medium and long, 72 / 140 / 220 at 2K

• Image set — the same per-image ladder applies to each frame in the series

Put next to that, "no published rate" is not a neutral absence. It means the Microsoft model cannot be costed at all, and a comparison of two image models in which one side of the subtraction is unknown is not a comparison of price.

A two-column comparison scoreboard titled 'Kling Image 4.0 vs MAI Image 2.7 Preview — the scoreboard', contrasting release date, editing Elo, text-to-image Elo, maximum resolution, published price and API access for the two models, with a footer reading 'Kling figures per Artificial Analysis; MAI Preview figures per Artificial Analysis as last recorded, row since withdrawn.'

The mask question, where both sides are thin

Mask-based editing — paint a region, change only that region — is the control that makes an editing pipeline reviewable. A reviewer can see the painted area and judge only that area, which is why production tools have shipped it for a decade. Microsoft's description of the Preview is instruction-driven throughout: restyle, relight, replace the text. There was no published mask parameter, no bounding box, no reference-image maximum.

Kling's release does not answer the question either. Its capability list includes "layered editing", glossed as "full-page view, layered editing", which is the closest thing to a control interface in the announcement. Whether that is a mask, a layer stack, or instruction-following dressed in layout language is not something the price sheet resolves, and treating it as a mask until you have tested it would be a mistake. For the Kling side of this pair, the mask question is open rather than answered — which is a different thing from answered badly.

If you have an hour and one of these budgets

The Preview gives you a free look at what Microsoft's editing research was pointing at in October, and that is genuinely worth an hour if your work is whole-image restyling or relighting. It will not tell you what the November model does, and it will not tell you what anything costs.

Comparing Kling Image 4.0 against the models you can already call is a different exercise, and it is the one that produces a decision. OrcaRouter does not currently serve Kling Image 4.0 or any other Kling image model — the Kling identifiers in our catalogue are video — so this piece is not going to suggest otherwise. What the catalogue does carry is a route to the image models we do host, from the GPT Image line through Imagen 4.0 to the Gemini image models, all on one key with provider list pricing passed through at 0% markup. That is the honest way to put Kling's 6.1-cent 2K still beside a model you can call in the same afternoon: send the same ten prompts down both paths, then read two invoices instead of two announcements.

What Kling is actually selling in this release is not subtlety. It is a tier nobody else in this pair offers — 8K, at 15.4 cents a frame at the entry rate — plus a documented workflow for tall e-commerce layouts that only accepts an image-to-image request and climbs from 60 to 200 units depending on length. If that describes your problem, the comparison that matters is against your own current cost per asset, not against a number from a board that has since deleted the entry.

Screenshot of the OrcaRouter model catalogue in English, headed '226 models, 17 providers, one API key, one bill', with the modality tabs showing Text 190, Image 12, Embeddings, Video and TTS counts, the model list with per-million-token input and output prices, and the prepaid credit plans.

The Preview's disappearance is the more instructive event. It is a reminder that a leaderboard position is a snapshot of a population that keeps moving, and that the only image models worth planning around are the ones with an endpoint, a price and a row you can still find next month.