Runway Aleph 2.0 and the end of the reshoot

The release that changed production budgets this year was not a model that generates footage. It was a model that edits footage you already shot. Runway Aleph 2.0 and Edit Studio arrived on 21 May 2026 and got a fraction of the attention paid to the generators, which is backwards, because this is the capability that touches the line items a client actually argues about.

Aleph 2.0 does in-context editing of existing footage, up to thirty seconds at 1080p. It transfers a look from a reference frame. One edit propagates across multiple shots and through scene changes, which is the part that matters on a real job, because a note from a client is almost never about one shot. Edit Studio previews a change as a still before you spend a render on it. Runway's announcement.

Why editing beats generating on a live job

Generative shots are easy to sell and hard to revise. Tim Katz at Knucklehead put the production reality in one line: "changing a shot late is really hard." His conclusion is the right one and it is not the comfortable one. AI has to be planned in from the start of a job rather than bolted on when the schedule slips. The full interview with Katz and Chris Hewitt.

Hewitt, creative director at Airhead, names the other half of it. Leave details unspecified and the model fills the gap with, in his words, "the AI slop we see all over social media." A model has no instinct for the thing you did not ask for. It will always produce something, and the something will be an average of everything it has seen unless a person decided otherwise.

Put those two observations together and you get the case for an editing model rather than a better generator. The failure mode of generative work is a shot that is 90% right with a client note attached to the other 10%. Before Aleph, that note meant regenerating and losing everything that was already approved, or taking it into compositing at a cost that erases the saving the client was promised. An editing pass on the footage you already have is the difference between a revision and a reshoot.

What you pay

Runway's tiers: Free with 125 credits, Standard at $15/month for 625, Pro at $35/month for 2250, Max at $95/month for 9500. Current pricing.

Standard

$15/month

Pro

$35/month

Max

$95/month

One detail decides the tier for anyone doing paid work. Max is the only tier with HDR, ProRes and image-sequence output. Everything below it gives you a compressed deliverable, which is fine for a social cutdown and useless if the work is going into a real grade alongside camera footage. If a colourist is involved, the subscription question is already answered, and the lower tiers are a false economy that shows up in the DI suite.

“An editing pass on the footage you already have is the difference between a revision and a reshoot.”

Runway also now resells rival models on its own platform, which changes how you think about the account. It is less a single model and more a room you work in, with the indemnity position on Enterprise extending to those hosted models too.

One piece of context on Runway's earlier benchmark claim, because it still circulates: Gen-4.5 launched on 1 December 2025 at the top of Artificial Analysis text-to-video with 1247 Elo. The research post. The board was reset in summer 2026, so that score cannot be compared with any current number. Anyone quoting it against a 2026 ranking is comparing two different scales.

The alternative

Luma's Ray3.2 Modify Video V2 does the same category of work: editing existing footage, up to twenty seconds at 1080p. Its distinguishing feature is performance tracking of skeletal posture, gestures and facial expression for up to eight faces, which is the feature to reach for when the note is about a performance rather than a look. It also outputs native HDR with 16-bit EXR export. Luma's announcement. Shorter clips, deeper performance control, and a delivery format that drops straight into a finishing pipeline.

The work that actually ships is hybrid

Every piece of AI-led commercial work we can point to this year depended on exactly this capability. Promise composited more than 500 AI shots behind a host filmed on a volume stage for CBS. Knucklehead and Airhead shot the car for real on MG Motors' "The Pentathlete" and generated the pentathlete vignettes around it. In neither case was the model making the film. It was making the parts of the film that would otherwise have been impossible to afford, and then somebody had to make those parts sit with footage shot on a camera.

David von Ritter at cavalcade, on the AI-led films for Swissquote, gave the honest version of the standard: "AI is like plastic surgery: it's good if you can't see it." The detail underneath that quote is the useful part. The bull in those films had to be rebuilt in 3D with substantial R&D so that it could perform physically complex actions and stay consistent across shots. The full piece, 15 June 2026. The model did not solve consistency. A 3D asset and a team of people solved consistency, and the model did the part that followed.

That is the shape of the budget now. Less money on acquisition, more on asset construction and on the editing passes that make generated and photographed material agree with each other.

You stop paying for failed renders

One quiet change ran across every vendor in 2026: previz moved inside the model. Seedance has a 3D white-box pass. Runway's Edit Studio shows a still before the render. Luma has Draft Mode. The same idea in three places, arriving within months of each other, which usually means the vendors were all hearing the same complaint.

The complaint was that the cost of a generative pipeline was never the successful shots. It was the twenty attempts that preceded each one, and the fact that nobody could tell a client in advance how many attempts a shot would take. Previewing a change as a still costs almost nothing and kills most of that waste.

Two things worth changing on the next job. First, book the editing model at the same time as the generator, in the same budget line, so that revision capacity exists before the first client review rather than being found afterwards. Second, if the work is going to a grade, buy the tier with ProRes and image-sequence output on day one. Both decisions cost very little at the start of a job and are expensive to retrofit in week five.