AI made Meta creative cheap. Picking the winner still costs.

In 2022, a UGC video for Meta cost about $2,700 to produce.

That is not from memory. That is Statista's benchmark: video-based sponsored posts on TikTok averaged $2,102, YouTube $2,741, Instagram Stories $2,784.^[3] One creator, one video, one Meta-ad-length asset. That was the industry norm.

And that number set the pace of media buying. Foxwell Digital's Meta creative framework from that era treated two to four new creatives per week as the expected output at $50,000 to $100,000 per month in ad spend.^[2] Two weeks of production could exceed a month of media budget. That was the physical constraint on how fast a media buyer could iterate.

Fast forward to 2026.

The same 10-second Meta ad video costs $0.80 to produce on Google Veo 3 Fast.^[4] Five dollars sixty on OpenAI Sora 2 Pro. Meta Advantage+ Creative bundles it into Ads Manager at zero incremental cost to the advertiser.^[6] The top-quartile Enterprise Meta advertiser now ships 54.64 new creatives per week per Motion's 2026 benchmark.^[1] That is fifteen to twenty-five times the 2022 pace, at roughly one-two-thousandth of the unit cost.

Meta's winner rate did not move. It is still about 5%.

Ship 54, 51 lose.

The bottleneck moved. It used to be making the ads. Now it is picking which of them to run.

How production got cheap

The 2022 baseline. Statista's benchmark for video-based sponsored posts that year put the per-post cost at $2,100 to $2,800 across major platforms.^[3] That was for one asset. Vidico's standing benchmark for commercial video production sat at $5,000 to $25,000 for standard shoots and $50,000 to $100,000 for high-end brand campaigns.^[10] A media buyer sourcing UGC from a creator marketplace paid $2,000 to $3,000 per usable video. A media buyer commissioning a brand-quality spot paid ten times that.

Foxwell Digital's framework from the same era captured the physical consequence. At $25,000 to $50,000 per month in Meta ad spend, expected new creative output was one per week. At $50,000 to $100,000 per month, two to four per week. Only above $100,000 per month did the framework treat ten to fifteen new creatives per week as normal.^[2] Production, not paid media, was the throttle.

Then generative AI closed the cost curve. Here is what a Meta-ad-length video creative costs in August 2026.

The unit cost of a Meta-ad-length video collapsed from $2,700 to under $1 in four years. That is not a marginal efficiency gain. That is two orders of magnitude. And it is not theoretical. It is showing up in the shipping data.

Motion's Creative Benchmarks 2026, drawn from 578,750 ads across 6,015 accounts and $1.29 billion in Meta spend, records enterprise-tier top-quartile weekly creative output at 54.64.^[1]Cross-referenced against Foxwell's 2022 pre-AI framework of two to four creatives per week at similar spend brackets, that is a fifteen-to-twenty-five-times lift in top-of-funnel volume.

Production went from $2,700 per unit to $1 per unit. Volume went from two per week to fifty-four per week. And the media budget to test each new creative did not move.

The cost curve, 2022 → 2026

Why Meta's winner rate did not move

The predictable response to a two-orders-of-magnitude cost collapse would be: winner rates rise, because every advertiser is now producing better creative faster. That is not what happened.

Meta's auction ranks ads relatively, not absolutely. It picks the creative most likely to earn the impression given every other creative competing for that impression at that moment. When production speed doubles industry-wide, every account's creative competes against every other account's creative at the new higher volume. The 5% winner rate is a distribution property of the auction, not of the creatives. It does not move because generation got cheaper. It only moves if the auction changes.

The adoption numbers back this up. Meta's own Q1 2026 earnings disclosed that 8 million advertisers were using at least one of Meta's AI creative tools, roughly doubling from 4 million at the end of Q4 2024.^[7] The IAB State of Data 2026 put creative-process AI deployment at 83% of ad executives, up from 60% in 2024.^[8] Every advertiser adopting AI production at the same time is exactly the condition under which winner rates stay flat. The auction rebalances at the new production tempo.

Which is a useful diagnostic. If your first thought was ‘so my AI-generated ads should be winning more,’ the correct update is: no, they should be losing at the same rate as everyone else's AI-generated ads. The individual advantage-of-cheaper-production is real. The industry advantage is not. The math of an efficient auction rules that out.

The bottleneck math

Now the specific numbers.

Motion's top-quartile Enterprise ship rate: 54.64 creatives per week.^[1]

WordStream's 2025 aggregate US CPL on Meta: $27.66.^[9] Meta's own Learning Phase threshold: roughly 50 optimization events per ad set. Cost to test one creative to that threshold: $1,383.

Multiplied out: 54.64 x $1,383 = $75,548 per week to Meta-test every creative you ship. That is $302,000 per month in test spend alone. That is a top-quartile Enterprise-tier monthly Meta budget applied entirely to Learning-Phase discovery, before a single dollar of scaled media.

Nobody does this. Nobody can afford to.

Which means the 54 candidates have to be filtered somewhere between production ($0.80) and Meta test ($1,383) before any of them enters the auction. That filter is the new bottleneck. It is not production. Production is solved. It is not attribution or measurement. Those are downstream. It is the pre-launch judgment call on which of the 54 deserves the $1,383 Meta test.

Why the existing testing tools were built for the old world

PickFu, Marpipe, Poll the People, Kantar, Nielsen. All designed when the media buyer's problem was ‘I have three to five finalists, help me pick.’ Their unit economics were built for that regime. PickFu at $15 to $1,000 per poll made sense when you ran three polls per campaign. Total pre-launch test spend, under $5,000. Cheaper than Meta by an order of magnitude.

Now the media buyer has 54 candidates. If PickFu costs $150 per creative to test, filtering 54 candidates costs $8,100 per week. That is $34,000 per month. That is not cheaper than Meta. That is comparable to a Meta test budget. The economics do not scale to AI production speed.

The tools were designed for a different volume regime. Fine tool, wrong-shaped market.

Poll the People at $15 to $500 per poll runs into the same math. Marpipe at four-figure monthly enterprise pricing charges at the production-cost tier a decade ago. Kantar and Nielsen were never priced for individual-creative testing; they are built for annual brand studies. syntheticusers.com adjacencies at $99 to $199 per study are closer, but per-study, not per-creative-at-scale.

Every one of them is priced for a world where the media buyer had five finalists. In 2026, the media buyer has fifty candidates. The tools that assumed the old volume regime cannot serve the new one at the price point the new one requires.

What the new-bottleneck tool has to look like

Two constraints, both non-negotiable.

Unit economics of pre-launch judgment have to match unit economics of AI production. If a creative costs $1 to produce and $1,383 to Meta-test, the pre-launch filter has to price somewhere between those two numbers with meaningful headroom. Testing every creative at $150 does not work because it eats the ad budget. Testing at $1 to $2 per call works because it lands inside the production-cost regime.

Volume ceiling has to match production speed. A tool built to run three tests per campaign is not the same category as a tool built to run 50 to 100 filter calls per week. Different product architecture, different pricing model, different volume-of-calls guarantee. The gap is not incremental.

Solve for both: subscription pricing, high monthly call ceiling, pre-launch judgment at production-tier unit cost. That is the specific product category the bottleneck shift created.

What Splitroom looks like at that math

Those numbers are not accidents. They are what the new-bottleneck economics demanded. Pre-launch judgment at the same unit cost as AI production is the specific category Splitroom occupies.

Static creatives specifically (see the companion piece ‘Meta's top-winning ad format is not video. It is text.’). Two go in. Up to a thousand synthetic buyers argue them out. Verdict, segment splits, dimension attribution in about nine minutes. Not ‘pick a winner from three finalists.’ Filter your 50 to 100 weekly candidates down to the three or four worth burning Meta spend on.

The old testing tools sold a compare-finalists workflow. That was the correct product for the old volume regime. The new-bottleneck tool sells a filter-candidates workflow. Different job. Different math.

The market moves. Tools follow.

For the first time in a decade, the pre-launch filter is the highest-leverage tool in a Meta media buyer's stack. Not production, which is solved. Not measurement, which is downstream of the auction. Not attribution, which is a scoreboard. The specific question of which of the 54 weekly candidates deserves the $1,383 Meta test is where marginal dollars now compound.

Every existing creative-testing tool priced above that math is going to have to reprice or disappear. The founders of those tools built for a real, correct 2018 problem. The problem moved. The tools have to move too, or new tools built inside the current bottleneck economics will take the market.

The old bottleneck was making the ads. That bottleneck is gone. AI killed it. Motion's data records the aftermath: 54.64 creatives per week at the top quartile, at $0.80 per creative unit cost, against a 5% winner rate that did not move.

The new bottleneck is picking which of them to run. That is the product Splitroom is built for. Priced for it. Volume-tiered for it. Category-locked for it.

Production got cheap. Judgment got expensive. The tools that recognize which bottleneck the market is actually solving get the decade.