Five Models, One Timeline: Adobe Turns Premiere Into the First Model-Agnostic AI Video Workbench

The most consequential AI video news of the week did not come from a model lab. It came from the timeline. On September 8, timed to IBC 2026 in Amsterdam, Adobe shipped the first multi-model, in-timeline generative AI workbench inside a major professional video editor: five competing video-generation models — Adobe Firefly, Google Veo, Runway, Kling, and Luma — now live side by side in Premiere Pro’s toolbar, generating footage directly into the cut without an export, a browser tab, or a re-import.

For two years, the “AI video workflow” for professional editors has been an oxymoron. You hit a gap in your B-roll, left your NLE for a standalone generation app, prompted, waited, downloaded, re-imported, and then discovered the clip was the wrong aspect ratio, the wrong frame rate, or simply the wrong visual register next to the surrounding footage. Every integration that promised to fix this left one of those steps intact. Adobe’s Generative Media Tool, shipping now as part of the IBC 2026 release, removes all of them at once.

How the Generative Media Tool works

The tool lives in the Premiere toolbar in the same slot as Generative Extend; a long press toggles between the two. An editor selects a range on any video or audio track, a generative taskbar appears alongside the range, and from there it is prompt, choose model, generate. The result arrives already trimmed to the selected range, at the sequence’s native resolution and aspect ratio, and lands in the project’s generation folder and a dedicated Generation History bin. Critically, the output behaves like any other clip — editable, colorable, nestable — not like a locked AI artifact.

The architectural detail that separates this from prior integrations is the reference-frame mechanism. An eyedropper tool lets the editor sample one or two frames directly from the timeline or the program monitor, and those frames travel to the model alongside the text prompt as image-conditioning inputs. Generated footage is therefore not conjured from scratch; it is conditioned on shots that already exist in your cut. Because every failed generation costs credits, anything that raises the first-attempt success rate — matching grade, framing, and lighting to the neighboring material — has direct financial significance for a working facility.

All five models are selectable from the same taskbar, with the credit cost for the chosen model and output settings displayed before the editor commits. Adobe has said the model list will expand as new providers are integrated. Only Firefly carries a publicly posted per-clip rate; the four partner models disclose their cost in the taskbar at the moment of selection, but there is no published rate card for them yet.

The credit math nobody is talking about

The pricing story buried under the feature announcements is worth pulling into the light. A five-second Firefly-generated clip costs 100 credits, a figure that holds across all four credit packs Adobe sells: 2,000 credits buy roughly 20 clips, 7,000 buy 70, 10,000 buy 100, and 50,000 buy 500. That works out to about $0.43–$0.50 per five-second clip depending on the pack.

Two numbers deserve attention. First, generated video at 20 credits per second costs four times more than audio translation (5 credits per second) in the same system. Second, the base Creative Cloud plan’s 2,000 monthly credits buy approximately one minute and 40 seconds of generated video. One five-second clip is 5% of a base-plan month. Any team that standardizes on generative B-roll needs to budget for credit packs, not just subscriptions.

There is also a genuine anomaly in the pack structure: the $49.99-per-month 10,000-credit tier costs more per credit ($5.00/thousand) than the $29.99-per-month 7,000-credit tier ($4.28/thousand) below it — the mid-tier plan is effectively 16.7% worse value per credit. The $199.99 50,000-credit plan (~$4.00/thousand) is the only tier cheaper per credit than the 7,000 pack. For teams doing the procurement math, the 10,000 tier is the worst purchase in the lineup, and two 7,000 subscriptions ($59.98 for 14,000 credits) beat it outright.

The strategic admission

The model-agnostic architecture is the part of this release that matters most for the industry, and it encodes a competitive reality Adobe has never stated directly: Firefly Video has fallen behind the field. Independent comparisons through 2026 consistently place Firefly Video behind Google Veo 3.1, Sora 2, and Kling 3.0 by a meaningful margin.

Adobe’s answer is not to chase the model race but to change what the competition is about. By making Veo, Runway, Kling, and Luma selectable from the same toolbar as Firefly — inside the timeline, not a separate app — the question shifts from “which model is technically best” to “which workflow makes any model easiest to use.” Adobe is betting the answer is an NLE, not a standalone generation app. Runway operates its own NLE-adjacent interface; Kling and Luma live in web apps. Pulling those APIs into Premiere removes the last reason for an editor to leave, and Adobe earns credit revenue regardless of which model the editor picks, because the credits are sold by Adobe even when generation runs on a partner’s infrastructure. Whoever owns the timeline owns the workflow — and Adobe just made its timeline the place where every video model competes for the same shot. A quiet side effect: Adobe now gathers usage data across all four rivals’ models, which is an advantage nobody is pricing in.

Audio gets the same treatment

The Generative Media Tool extends to sound. Generate Sound Effects ships now: draw a range on an audio track, write a prompt, optionally record your own voice as a timing and intensity reference (the recording guides pacing, it is not the output), and the tool returns four saved variations to audition.

Two more audio features arrive in beta. Generate Music uses Adobe’s Firefly Music model to produce original compositions with tempo and looping controls, and Adobe states all music produced through it is commercially licensable with synchronization rights across media types — for commercial work, where an unlicensed track in a deliverable is a legal liability, this is the detail that matters. Generate Soundscape takes a different approach entirely: it analyzes up to 15 seconds of video above the selected range via computer vision, needs no prompt at all, and returns four pre-mixed tracks — two ambient, two sound effects — each with four variations. That is 16 audio assets from a single gesture. Hands-on testers report mixed results, with some tracks nicely usable and others out of sync.

A separate beta track expands audio repair. Enhance Audio splits a mixed recording into dialogue, music, ambience, and effects with independent sliders. Separate Crosstalk divides two overlapping speakers into independent tracks plus a residual — an answer to one of the most persistent pain points in interview and documentary editing. Dynamic Auto Ducking writes music-under-dialogue keyframes automatically and regenerates them whenever clips move. None of the three carries a stated credit cost.

After Effects gets an operator, not a generator

The After Effects AI Assistant, in public beta, is a different species of feature: it does not generate media, it operates the application. It works across the entire project rather than the open composition, and its two flagship capabilities target the two most dreaded tasks in motion work. First, expressions: describe what you want in plain language and the assistant writes the JavaScript, builds the rig with adjustable slider controls, shows its reasoning step by step, and fixes broken rigs on request instead of forcing manual debugging. Second, organization: it can review a project with thousands of layers, propose a restructuring plan, and execute it after approval — a lifeline for anyone who has inherited an undocumented project file. Every action uses native After Effects operations and is therefore undoable. In Adobe’s press demonstration, a stop-motion collage rig with expression-driven controls across six pre-comps was built from a single prompt.

An Emmy, a release, and a hardware cliff

The IBC announcement lands within six days of the Television Academy announcing that Premiere will receive an Engineering, Science and Technology Emmy on October 14 — the first for Premiere specifically, honoring 35 years of the application and citing its machine-learning and generative AI functionality as native timeline operations. The release also widens Color Mode’s beta with Auto Color, HSL and Luminance masks, and new Style Modules; redesigns the Properties and Effects panels; and opens Frame.io Mounted Storage, which streams full-resolution cloud media directly into the timeline without local download, to all Creative Cloud accounts.

The context is a moment of transition: Version 27 later this year will require Apple Silicon Macs, cutting Intel-based machines off from new features. The bets Adobe is making here — model-agnosticism over model supremacy, workflow ownership over generation quality — are how a 35-year-old application plans to stay the center of professional video while the ground underneath it generates new footage one prompt at a time.

What to watch: whether partner-model rate cards get published once IBC reactions land, whether the model list grows past five, and whether the Generate Music sync-rights claim survives contact with broadcast legal departments. If Premiere has genuinely become the routing layer for every video model, the next fight is over who sets the price per shot.