The model war is over.

Tool choice is no longer the interesting question.

Working independently, the three panels settled the debate the discourse is still having. That convergence is itself the finding: when thirty professionals in three different jobs assemble the same spine, the tools have become infrastructure, and the attention of the people who ship work has already moved to what infrastructure cannot decide. The rest of this edition is about where it went.

Sources: all three issues, stack sections.

Just as telling: what thirty stacks do not contain.

Not one contributor names Sora in the stack they ship with. Adobe's editors sit at the end of most of these pipelines, yet Firefly, Adobe's own generator, is not mentioned once by anyone. And for all the open-source noise, only five of the thirty go anywhere near open-weight pipelines, four of them filmmakers at the technical frontier, all through ComfyUI-style node tools. Absence at this scale, across three independent panels, is data: launch hype and working adoption have fully decoupled.

Sources: stack walkthroughs, all three issues.

Generation changed everything. The cut changed nothing.

Twenty-nine of the thirty full pipelines described across the series end in a conventional editor: Premiere, Resolve, CapCut. (The single exception finishes in platform-native editors.) Two years into the generative boom, the timeline is untouched, and several contributors argue post-production is precisely where AI-generated work stops looking AI-generated: pacing, grading and sound are where the seams disappear, and only the edit controls them.

Sources: stack walkthroughs, all three issues.

Claude fronts the pipeline. ElevenLabs owns the voice layer.

Claude appears in 25 of 30 stacks as the thinking layer: research, scripts, shot lists, prompt systems, and increasingly workflow automation. And across all three panels, ElevenLabs is the only voice tool named more than once, the quietest near-monopoly in the whole dataset. (Disclosure: Vibedex uses Anthropic models as benchmark judges; the count is the panels' own.)

Sources: stack sections, all three issues.

The image layer is vertical. The video layer is not.

Same tools, same summer, three different image-layer winners: Nano Banana for filmmakers (asset volume), GPT Image 2 and Nano Banana for ad studios (label and product accuracy), Midjourney for social creators (aesthetics). Meanwhile Kling and Seedance carry motion in every vertical, and the only split is philosophical: craft verticals optimize the best case (Seedance), commercial verticals optimize the worst case (Kling). If you benchmark or buy image models without naming the job first, you are measuring the wrong thing.

The stable video layer also carries a geographic fact our filmmaking panel spotted first and the full dataset confirms: the working motion layer is overwhelmingly Chinese. Kling (Kuaishou), Seedance and Dreamina (ByteDance), PixVerse and Hailuo, against an American minority of Veo, Runway and Luma. The words are American, the motion is Chinese, and one company sits at three layers of the creator stack: ByteDance, via Seedance, Dreamina and CapCut.

See the matrix above; per-issue image sections.

The AI-disclosure gradient: who you hide the AI from depends on who judges you.

Filmmakers lead with AI; it wins festivals. Ad studios hide it from clients; the moment a client thinks "AI-generated," they start looking for flaws instead of results. Social's best performer hid it from the audience; the campaign worked because AI supported the story instead of being the story. One law underneath: AI performs best when the evaluator is not asked to admire it.

Sources: Ads s04, Social s05, Filmmaking passim.

Reputation lags reality, in both directions.

Once the job, not the tool, is the unit of analysis, the market's pricing of reputation starts to look wrong. The clearest cases in the data:

The pattern: tools are priced on reputation, but practitioners pay for workflow fit. When those diverge, the reputation is what's mispriced.

Sources: overrated and verdict answers across all three panels.

AI's real economic gift is iteration, and each field spends it differently.

What the savings bought differs by vertical. Filmmakers bought access: films and worlds that were locked behind impossible budgets. Ad studios bought testing velocity: from a handful of ads a month to 10 to 20 tested variations a week, with one solo creator shipping 25 to 30 product videos monthly. Social creators bought scale: one contributor went from zero to 3M+ subscribers and a billion views on a daily schedule that could not exist without AI at every stage.

Read those three purchases together and a bigger shift appears: the unit of production collapsed to one person. A solo creator now ships what used to take a studio day per asset; a solo channel now operates at audience scales that used to require a media company. AI did not make studios cheaper. It made individuals into studios.

Sources: economics sections, all three issues.

The work did not shrink. It moved upstream.

Ask where the panels spend the hours the tools freed and the answer is unanimous: before generation. Reference sheets locked before any motion touches a character. Character passports attached to every generation. Shot lists written like a director on set, lens and light called out loud. One studio head's discipline stands for the whole cohort: define what must remain invariant before generating anything, then change one variable per batch.

The reason is structural. Generation is cheap and unpredictable, so control migrated to the only place it still pays: the plan. Pre-production absorbed the saved time, and it is now where the craft lives. The tools compete furiously at the generation step; almost nothing serves the planning step where these professionals actually spend their day.

Sources: pro-tip and workflow sections, all three issues.

Image-first is a production law.

The upstream discipline has one universal rule. Do images before video. Lock references before motion. Never go text-to-video for commercial work. A weak source image makes a weak video regardless of the video model. Every panel arrived at this rule in its own words, from its own economics: stills are where control is cheap, motion is where mistakes get expensive.

Sources: Filmmaking (beginner consensus), Ads (three independent phrasings), Social (craft school).

The unit of value is shifting from the clip to the character.

Asked where the field goes next, contributor after contributor described the same object: not better clips, reusable characters. A creative director building owned visual worlds and reusable product and character models. An ad creator predicting brands will hold libraries of characters, environments and voices, reused across dozens of ads. A commercial video maker building brand characters she expects audiences to recognize the way they recognize logos. A social creator betting that human influencer economics change permanently the moment any tool cracks cross-session character consistency.

The asset stops being the video and becomes the thing that generates videos. The one live disagreement is ownership: the ads panel expects brands to own these personas, film and social expect creators to. Watch where the audience relationship lives; the persona will land there.

Sources: future-bets sections, all three issues.

Model quality became table stakes. The real buying decision moved to the platform layer.

Two panels independently produced anonymous critiques with the same shape: many platforms are "wrappers around the same underlying models," charging "a premium for convenience." Nobody disputed where raw quality originates. But read the thirty stacks closely and the practical conclusion inverts: because the frontier models are reachable through many doors, the decision that actually separates working stacks is which door. The panels describe five distinct platform archetypes, each matched to a different job:

Convenience, workflow fit, consistency tooling, credit pricing and trust are what platforms actually compete on, and they are the least-measured layer in the market. It is where our comparison work now concentrates: we already evaluate Creative AI Platforms by capability, workflow and trust, and that coverage is expanding.

Sources: Ads s08-s09, Social s07, platform splits across all issues.

The interface is dissolving into agents.

No survey question asked about it, and it surfaced at the technical edge of every panel anyway. Filmmaking's frontier runs node pipelines. Five of eleven ad stacks are quietly agent-mediated: a visual-canvas agent, generation driven from inside a Claude conversation, agent-directed all-in-one tools, custom API pipelines. Social's most sustainable single stack is an end-to-end agent. The direction is consistent: from an app you open to an agent you brief. Watch this one compound.

Sources: Ads s02, Social s06, Filmmaking maturity curve.

Consistency is the industry's #1 unmet need.

Asked what no tool does well, every panel gave the same answer without seeing each other's: holding a character, a world and a creative language steady across shots, scenes and sessions. The requested product is remarkably specific and remarkably consistent: a persistent creative-director layer that remembers everything so the creator doesn't have to. Define a character and world once; direct scene by scene; the tool holds continuity.

Whoever ships this owns the next phase of the market. The panels have effectively written the spec.

Sources: all three issues, gap sections.

Taste is the bottleneck now, and the pros know it.

Asked what is most overrated, the social panel's near-consensus pointed past tools entirely, at the belief that the newest model makes better work. The filmmakers' version: they are filmmakers who use AI, not AI filmmakers. The ad studios' version: the AI does the work, the human owns the taste. Production stopped being scarce; judgment didn't.

Sources: overrated sections, all three issues.