The Death Announcement Is Wrong. The Death Is Real.
The public narrative around Sora 2 has been binary. Either "AI will destroy video production" or "it's still a toy — real production is safe." Neither matches what's actually showing up in client budgets.
What's showing up is a split. Premium production (six-figure brand films, retainer-tier content systems) is growing. Low-end production (explainers, UGC-style ads, social cutdowns) is getting eaten. And the middle — the $5K–$25K branded video tier that was the bread and butter of most regional agencies — is collapsing.
Our internal book of business through Q1 2026 showed a 41% YoY increase in retainer revenue and a 68% decrease in sub-$25K project inquiries. Every mid-tier agency we compare notes with is seeing the same curve.
What the Mid-Tier Actually Sold
The mid-tier agency model was built on a single value proposition: "good enough production at a reasonable price, without the overhead of a top-10 shop." For fifteen years that was a durable business. Marketing teams needed recruiting videos, product explainers, event recaps, and a steady drip of social content, and the mid-tier agency was the obvious answer.
What Sora 2 — and Veo 3, and Runway Act-Two, and whatever ships next month — actually compete with is that "good enough" tier. Not the premium work. The model doesn't beat a cinematographer who knows how to light a face. It beats a junior editor turning around a 45-second motion graphic for a LinkedIn post.
And that's 60–70% of what a mid-tier agency bills.
The Data Nobody Likes to Publish
The numbers most mid-tier shops won't put on their website:
Sub-$10K project inquiries are down roughly 65% across the regional B2B agency landscape in the first 15 months of generative video being credibly usable. $10–30K projects are down around 35%. Premium ($75K+) and retainer work ($12K/mo+) is up.
That's not a soft market. That's a structural shift in where the demand sits. Marketing teams who used to buy 8 mid-tier videos a year are now buying 2 premium ones and filling the gap with internal AI output.
The Counter-Argument, Steelmanned
Here's the strongest case against this framing: AI output still looks uncanny, it can't do brand consistency, and enterprise legal teams are blocking it. All true. And all of it buys the mid-tier maybe 18 more months.
The uncanny valley is narrowing. Brand consistency through fine-tuned models is an active product roadmap at every major vendor. Legal acceptance is a lagging indicator that follows vendor indemnification — and those indemnification clauses are getting written now.
The mid-tier agency that bets on "AI can't do it yet" is making a 2023 bet in a 2026 market.
What's Actually Growing
Two things. First, premium craft — the top 10% of production work, where cinematography, direction, and on-camera talent create measurable differentiation. That work is growing because it's the only thing that can't be faked.
Second, production systems — retainer-structured content operations that combine direction, capture, post, and distribution into a repeatable machine. The systems layer grew specifically because marketing teams realized they could run AI tools in-house, but they couldn't build or run the system around them.
Both of those are the opposite of mid-tier. One is craft you can't replicate. The other is operational infrastructure you don't want to build in-house.
What To Do If You're Buying Video
If you're a marketing leader looking at 2026 video budgets, three moves make the math work:
Stop buying mid-tier "good enough" production. You'll either regret paying for it in six months, or you'll find out your internal AI stack was always enough. Either way, the money's wasted.
Reserve premium budget for the 3–4 assets a year where craft is actually load-bearing — brand films, customer stories where trust is make-or-break, founder-led content where face and voice are the product.
Move everything else to a retainer-structured content system. Not a "subscription video" package. A system with cadence, measurement, and a specialist team. The math works even with AI tools in the stack, because the system value is the operation — not the shots.