What crossed the line in 2026

In November 2025 the working ceiling on a generated shot was five to ten seconds, at 720p or an upscale pretending otherwise, with the audio added afterwards and the sync fixed by hand. Everything about a commercial pipeline followed from that: you cut around the limit, you budgeted for failed renders, and you treated AI as a source of inserts rather than scenes. Six of those constraints are gone. This is the list for anyone who stopped watching twelve months ago, which is most clients.

Thirty seconds in one pass

Four models now generate thirty seconds in a single pass: Seedance 2.5 in July 2026, Wan 3.0 in August, Kling 4.0 in October, and Runway's Aleph 2.0 in May for editing rather than generation. Thirty seconds is not an incremental gain over ten. It is the length of a television spot. A shot that runs the full duration of a scene removes the stitching work, and more importantly it removes the creative compromise of writing to a five-second unit. The cut can go where the edit wants it.

Native 4K at 10-bit

Seedance 2.5 generates native 4K at 10-bit, and ByteDance is explicit that it is not upscaled. The distinction is worth pressing on because the alternative is still common. Google's Gemini Omni 1.1 Flash lists 1080p and 4K, and both are upscales from a lower native render. An upscale survives a web player. It does not survive a grade, a large screen, or a client's post supervisor pulling the file apart. If the deliverable is broadcast or cinema, native and upscaled are two different products wearing the same number.

Audio generated with the picture, not after it

Seedance 2.5, Wan 3.0 and MiniMax H3 generate audio in the same latent space as the image. Sync stops being a post-production task because it was never separate. Last year the sequence was generate, then score, then line it up, then regenerate when the lip movement refused to agree with the line. Now a clip arrives with its own footsteps, room tone and dialogue already attached to the frames that caused them. The saving is not the audio, which still gets replaced on most jobs. It is the elimination of a loop that used to run several times per shot.

Reference counts went up an order of magnitude

Seedance 2.5 accepts up to fifty multimodal references, against twelve in its previous version, and the types have widened: images, audio clips, 3D white models, style references. Fifty references is a different kind of instruction from twelve. It is closer to handing over a lookbook, a location scout, a wardrobe rail and a reel than to describing a shot. For brand work with a locked identity, this is the mechanism that makes consistency a setting rather than a negotiation.

Editing real footage, at a usable length

Runway's Aleph 2.0 and Edit Studio, from May 2026, do in-context editing of existing footage up to thirty seconds at 1080p, with one instruction propagating across multiple shots and through scene changes. Note what that sentence does not contain: generation. This is AI applied to material someone shot, which is most of the work in most campaigns. Change the time of day across a sequence, replace a product, carry a look from a reference frame onto footage that was lit differently. The propagation across scene changes is the part that was not possible a year ago, when every shot had to be treated individually and nothing matched.

Previz moved inside the model, and this is the one that changes a budget

Seedance has a 3D white-box previz pass. Runway's Edit Studio gives a still preview before you commit a render. Luma has Draft Mode. Three vendors arriving at the same feature in the same year is usually a sign that the problem was real.

Here is why it matters more than resolution. The hidden line in every AI production budget has been failed generations. You pay for renders that come back with the wrong hands, the wrong camera move, a product label that drifts, a performance that misses. Nobody itemises it for a client, and it is the largest variable in the estimate. Previz converts that spend into a decision taken before the money is committed. You approve a camera move as a white-box pass. You approve a frame as a still. Then you render once, knowing what you are getting.

“The most visible AI commercial of the year took a conventional commercial's timeline.”

For a client, the practical effect is that the unpredictable part of the estimate shrinks. The contingency on an AI job has always been about render volume rather than weather or crew overtime. That contingency is now partly a craft decision instead of a gamble, and a studio should be able to show you the previz stage as a line on the schedule.

Open weights reached the frontier

In 2025 open models sat a clear tier below the commercial ones, and the gap was the reason nobody built a serious pipeline on them. MiniMax H3 now ranks in the top three for text-to-video on independent blind preference voting. A post-train of it by fal topped the board. Pika shipped a commercial product, H3 Flex, built on it. The consequence for brand work is less about cost than about control: weights you can host are weights that cannot be switched off by a vendor announcement.

Which is the seventh crossing, and the one nobody predicted. A major lab retreated. OpenAI shut Sora down, the app in April 2026 and the API in September. Availability is now a variable in this category, not an assumption, and it belongs on the same list as any capability gain.

What did not cross

The gap between these thresholds and a finished campaign is still the work. Svedka's "Shake Your Bots Off" ran at Super Bowl LX as the first primarily AI-generated national Super Bowl spot, made with Silverside AI. Sazerac's account of it is the useful part: about four months, and humans wrote the story. The most visible AI commercial of the year took a conventional commercial's timeline.

Several of the loudest claims also have nothing behind them. There is no Veo 4, whatever a search result tells you. Kling 4.0's 4K and 10-bit HDR are listed as coming soon and have not shipped. ByteDance's figure for improved prompt adherence is its own, with no independent test. And the cost spread is wide enough to decide a pipeline on its own: Seedance 2.5 runs roughly seven times the per-minute cost of MiniMax H3.

Two questions settle most of this when you next brief a studio. Ask which model is attached to which shot, and why that one. Then ask to see the previz stage as a dated line on the schedule, with an approval against it. A studio working to the 2026 thresholds can answer both in a treatment. A studio still working to last year's can answer neither, and you will find out at the render stage, which is the expensive place to find out.