Vidu Q4 Preview vs Kling 3: Four Kling Variants, One Vidu Build, and a 124-Point Gap

The Artificial Analysis image-to-video board carries four separate Kling 3 scores and one Vidu Q4 Preview. The four Kling entries — Kling 3.0 1080p (Pro), Kling 3.0 720p (Standard), Kling 3.0 Omni 1080p (Pro) and Kling 3.0 Omni 720p (Standard) — occupy 1,055, 1,051, 1,044 and 1,036, a nineteen-point spread. Vidu Q4 Preview is at 1,179. It is a single model with a single score, and it sits 124 points above the best Kling 3 variant on that board, which is a wider gap than Kling's entire internal range multiplied by six.

That framing is worth stating at the top because most "Vidu vs Kling" pages compare one name against one name, and Kling 3 is not one name. Vidu Q4 Preview shipped 7 October 2026 as Shengshu Technology's first public preview of its next-generation flagship, ahead of the full Q4 model, with two modes: Image-to-Video and Reference-to-Video. Kling 3 arrived in February 2026 from Kuaishou as a tiered family that has since grown an Omni branch, a Turbo route and a Standard/Pro split at two resolutions. Comparing "Kling 3" to anything means picking which of those you mean.

The board, with every Kling variant named

Artificial Analysis, read 8 October 2026, audio kept in, AA-Video-I2V v1.0:

• Vidu Q4 Preview — 3rd, 1,179 ±10, 5,543 samples, Oct 2026, $7.20/min

• Kling 3.0 1080p (Pro) — 20th, 1,055 ±6, 14,896 samples, Feb 2026, $20.16/min

• Kling 3.0 720p (Standard) — 21st, 1,051 ±6, 14,692 samples, Feb 2026, $15.60/min

• Kling 3.0 Omni 1080p (Pro) — 22nd, 1,044 ±7, 6,185 samples, Feb 2026, $16.80/min

• Kling 3.0 Omni 720p (Standard) — 24th, 1,036 ±7, 6,198 samples, Feb 2026, $13.44/min

And on text-to-video (AA-Video-T2V v2.0), a separate board with its own scale and its own vote pool:

• Kling 3.0 Omni 1080p (Pro) — 15th, 1,015 ±10, 3,778 samples, Feb 2026, $8.40/min

• Kling 3.0 1080p (Pro) — 16th, 1,000 (anchor), 5,677 samples, Feb 2026, $10.08/min

• Vidu Q4 Preview — not present

The anchor note matters. Kling 3.0 1080p (Pro) is the reference point for the text-to-video board's Elo scale — the model every other entry is calibrated against, fixed at 1,000 by construction. That is a design decision by Artificial Analysis, not a score Kling earned, and quoting it as "Kling scored 1,000" without the anchor qualifier would be wrong.

Read the image-to-video block with the intervals in mind and one conclusion is unavoidable: the four Kling variants are statistically indistinguishable from each other, spanning nineteen Elo across a $6.72 per minute price range. Paying for Pro over Standard, or 1080p over 720p, buys you four to eight Elo. Paying for the Omni branch buys you nothing at all on this board — Omni 1080p Pro scores eleven points below the plain 3.0 1080p Pro, well inside the combined interval, at a lower list rate.

The comparison that survives scrutiny

Vidu Q4 Preview versus Kling 3.0 1080p (Pro) is the matchup worth running, because it pairs the best-scoring variant of each family on the board where both appear.

• Elo — Vidu Q4 Preview 1,179 ±10 over 5,543 votes vs Kling 3.0 1080p (Pro) 1,055 ±6 over 14,896 votes, a 124-point gap

• Published per-minute rate — Vidu Q4 Preview $7.20 vs Kling 3.0 1080p (Pro) $20.16, as recorded by Artificial Analysis

• Vote maturity — Kling's number rests on nearly three times the samples and is eight months older, which cuts both ways: better measured, and about a different competitive field

• Maximum resolution — Vidu Q4 Preview up to 4K with 2K and 4K at 10-bit colour vs Kling 3.0 up to native 4K on the OrcaRouter model card

• Clip length — Vidu Q4 Preview 3-16s on Image-to-Video, 1-16s on Reference-to-Video vs Kling 3.0 3-15s, up to six shots in a single generation on the Turbo route

• Reference inputs — Vidu Q4 Preview up to 15 reference images and up to 3 reference audio clips vs Kling 3.0's subject and motion control, not published as a numeric reference budget for image-to-video

A 124-point gap is roughly eight times the width of the two intervals combined. It is not a close call on that board and it would not become one if the samples were equalised. What it is, is a comparison between a February 2026 tiered family and an October 2026 preview from a smaller lab — which is the honest description and also the point.

The pricing comparison you cannot actually run

Here is where a straight rate line would mislead, so it is worth spelling out instead.

The $20.16 and $7.20 figures above are the per-minute API rates Artificial Analysis records for each model. They are useful as a like-for-like reference. They are not the price you pay if you call the model through a router, because routers price video per call or per second depending on the route, and a per-call price divided by an unknown clip length is not a per-minute rate.

OrcaRouter's Kling routes illustrate the point. kling/kling-v3 is listed on our own model page at a flat per-call rate of $0.084, with the page's own note that flat per-call pricing is how the image-generation and video routes are metered. Divide that by a five-second clip and you land near $1.01 per minute; divide it by fifteen and you land near $0.34. Neither figure is the $20.16 on the Artificial Analysis board, and neither is wrong — they are describing different things: a list rate the vendor publishes versus what a particular route charges for one generation at a particular length. kling/kling-3-turbo and kling/kling-v2-6 carry their own per-call rates as well, at $0.112 and $0.042.

The practical consequence: if you are budgeting for Kling 3, the number that matters is your clip length multiplied by the rate on the route you actually call, and if you are budgeting for Vidu Q4 Preview, the number that matters is the same arithmetic against a preview price Shengshu explicitly calls promotional. Compare lists to lists and invoices to invoices, and do not cross the two.

Where Kling 3 is the only answer

Two things Kling 3 does that Vidu Q4 Preview does not, and neither is a small one.

Multi-shot generation. The Kling 3.0 Turbo route on OrcaRouter is documented as producing up to six shots in a single generation, which is a different kind of tool from a model that renders one continuous clip. Vidu Q4 Preview's 16-second ceiling and reference-to-video mode are aimed at keeping one scene consistent; Kling's shot list is aimed at producing a sequence. For storyboarded work, that is not a marginal difference.

Text-to-video. Vidu Q4 Preview's product page lists two modes and its API documents two endpoints — /ent/v2/img2video and /ent/v2/reference2video — with no text-to-video path anywhere. Kling 3.0 1080p (Pro) is the anchor model on the text-to-video board. If your input is a prompt rather than a frame, one of these two models is simply unavailable to you, and no Elo comparison changes that.

Worth adding on our own side: we route Kling. kling/kling-v3, kling/kling-3-turbo and kling/kling-v2-6 are live on the public model API and callable on the same OpenAI-compatible endpoint as the text models, at provider list prices passed through with no markup added, so a vendor rate change lands the same day rather than at renewal. Vidu Q4 Preview is not an OrcaRouter route, and this page should not be read as saying it is — the video models we serve are on one key and endpoint alongside everything else, and adding a preview build when it goes GA is a routing decision rather than a rebuild.

Which Kling, and whether to bother

If you are choosing within the Kling family, the board says what the marketing does not: the four image-to-video variants are one score wearing four price tags. Kling 3.0 720p (Standard) at 1,051 is four Elo below 1080p (Pro) at 1,055 and $4.56 per minute cheaper on the recorded rate. The Omni branch does not earn its premium on this board at all.

If you are choosing between families: on image-to-video, Vidu Q4 Preview is 124 Elo above the best Kling 3 variant at roughly a third of the recorded per-minute rate, and that is the finding. But it is one preview build against a production family, it has no text-to-video path, and its price is promotional. Kling 3 answers more questions and scores lower on the one board where they can both be measured.

What would change this: the full Vidu Q4 release, which by Shengshu's own account is where the feature set settles — if it adds text-to-video, the overlap stops being a single board. And AA-Video-I2V v2.0, announced as coming with a revised taxonomy and 1080p throughout, which would replace these votes rather than re-sort them. Until one of those lands, the number to keep is 124, dated, with the board name beside it.