AI Product Video Cost: The $2,000 NBA Finals Ad Took 400 Tries

In June 2025, viewers of the NBA Finals saw a commercial with a shirtless old man draped in an American flag, a farmer floating in an inflatable pool full of eggs, an alien chugging a pitcher of beer, and a woman in a sparkly pink tracksuit driving a Zamboni. It was an ad for Kalshi, the prediction market, and none of it was filmed. Every shot came from Google's Veo 3.

The filmmaker, PJ Ace, said it took $2,000 and two days . Kalshi said it had tried production companies first and been quoted "in the six or seven figure range." A primetime ad made for the price of a used laptop made headlines.

The detail that got less attention was in Ace's own description of the work: it took him "300–400 generations to get 15 usable clips."

That's 20 to 27 tries for every clip that made the cut. The $2,000 was real, but most of it paid for footage nobody saw. That ratio, more than the price per second, decides AI product video cost. With the holiday ad season starting, it's worth working out before you spend.

The biggest ad season of the year

The National Retail Federation expects US holiday sales to pass $1 trillion  for the first time this year, up 4.2% to $1.013 trillion across November and December. It also expects shoppers to be careful, with its chief economist warning that "inflation, rising gas prices and pressure on household spending will affect consumers differently across income groups." Careful shoppers take more convincing, so stores need more ads, and more versions of each one.

That's the part AI video changes for a small brand. Ad platforms learn faster from a campaign with many variations: different hooks, backgrounds, first three seconds. A brand used to have to choose between one polished spot and a dozen cheap ones. Ace predicted where this goes: "small teams making viral, brand-adjacent content weekly, getting 80 to 90 percent of the results for way less."

The question is what "way less" means once retries are counted.

Pricing 100 Black Friday ads

We priced a typical holiday push for an online store: 100 vertical product videos, 15 seconds each, so 1,500 seconds of finished video.

Two assumptions make a product ad much easier than the Kalshi spot:

• Start from your real product photo. OceanToken's video endpoint accepts a first frame on models that support it, so the product in the ad is the product you ship. A product invented from a text prompt invites returns and complaints.
• Keep the scene simple. One product, one setting, no crowd of characters. The Kalshi ad was deliberately outrageous, which is the hardest kind of shot to get right. A clean product shot fails much less often. We assume three tries per usable clip, so 4,500 seconds generated.

At each model's starting price per second (lowest resolution and mode; higher settings cost more):

Scripts and voice-overs barely register. All 100 scripts on anthropic/claude-sonnet-5.5, at about 1,000 input and 250 output tokens each, come to about $0.41, and 100 narrations of about 250 characters on fish-audio/s2-pro cost about $0.38. As with AI short drama, video is nearly the whole bill.

What retries do to the bill

Three tries is our estimate for simple product shots. If your scenes look more like Kalshi's, the bill changes fast:

Going from three tries to twenty multiplies the cost by almost seven on either model. No change of model saves as much as making each shot easier to get right: start from a real photo, use one subject, keep the motion simple, and write the prompt carefully before you hit generate.

The plan worth copying

The cheapest way to make all 100 ads is the budget model at $193.69. But then your best ad also runs at budget quality for the whole season.

The last row of the first table costs about the same and follows the way performance marketers already work:

1. Make all 100 on a fast, cheap model. Two tries each, $129.12. These are tests.
2. Run them for a few days. Let the ad platform show you which hooks and angles get clicks.
3. Re-render the top 20 on a premium model. Same script, same first frame, better motion and lighting. 300 seconds on Veo 3.1 is $54.00.
4. Put the ad budget behind the re-renders.

You spend about what the budget plan costs, and the ads carrying real money are premium ones. Making all 100 on the premium model would cost $810.00, mostly on versions nobody clicks.

Before you start

• Resolution and length. These are starting prices. 1080p or longer cuts raise every row, roughly in proportion, so the comparison holds.
• Measure your own tries. After the first 20 ads, count how many generations each usable clip took. That number matters more than any price in this article.
• Time. Video jobs take minutes, not seconds. Generating 4,500 seconds the weekend before Black Friday is cutting it close. Start in the first half of November.

Doing it in one place

Every model above is on OceanToken behind one API key and one balance, at the listed price with no platform fee on top, so you can draft on Seedance and re-render on Veo without opening a second account. Video jobs hold their estimated cost when they start and are charged the actual cost when they finish, and failed jobs aren't charged. If you work in Claude Code or Codex, the OceanToken tools can estimate a batch before you run it.

Kalshi got a primetime ad for $2,000 and 400 tries. With a real product photo, simple shots and a plan for which ads deserve premium renders, you can make a hundred for less.

See every video model and price on OceanToken →

Prices are OceanToken starting prices from the model catalog as of November 5, 2026, at each model's starting resolution, and every figure worked out from them follows the catalog whenever this page is rebuilt. Tries per clip are estimates, except the Kalshi figures, which come from the filmmaker's own account as reported by Mashable.

Related: What an AI drama really costs, and how to pick the models