Ten Keeper Assets, Eight Tools, and a Spreadsheet Full of 'Almost'

Here is the least glamorous thing we did this month: we built a spreadsheet for creative failures. Not the fun kind of failure—the kind where a video model gives you 90 percent of a product shot and then hallucinates a second label. We wanted to know what it costs to get from 'almost' to 'ship it.'

The Brief and the Rules

On 2 November 2026, we sent the same brief to eight tools we labeled A through H. The brief was for a fictional US direct-to-consumer cooler brand: 10 publishable assets total. Four lifestyle stills, two product-in-context images, three short vertical video clips for paid social, and one six-second hero loop for a landing page. Every tool got the same brand guide, color palette, product reference photos, and a one-page acceptance checklist. No tool got extra prompting help beyond its own docs. We capped each run at 90 minutes of active human time and logged every generation attempt.

Usable meant: brand-safe, no garbled text, no extra product features, correct logo placement, no weird anatomy, and ready for a 1080x1920 or 16:9 crop. We did not count an asset as usable if it needed another model pass. That rule alone killed a lot of cheerful estimates.

The Metric That Made Us Uncomfortable

Cost per usable asset, or CPUA, is simple: tool spend plus human edit time plus licensing review overhead, divided by assets that actually passed. We used a blended US freelance rate of $75 per hour for edit and prompt work. Licensing review got a flat $40 per tool because someone on our team had to read the terms, check commercial rights, and confirm whether outputs could be used in paid ads. We did not include subscription waste from previous months; we only counted spend tied to this brief.

CPUA = (tool spend + labor + license review) / usable assets.

Once we ran the numbers, the advertised monthly price stopped mattering. The useful question was: how much does this tool cost before it produces something a brand manager will approve?

Where the Money Actually Went

Two patterns showed up fast. First, video tools burned credits in a way stills tools did not. Tool H spent $260 and still only got six publishable assets, mostly because every camera move introduced a new product shape. Tool F was cheaper per generation but generated so many near-misses that the edit stack became the real cost. A 20-second retry feels free until you pay someone to review it.

Second, the most expensive failures were not ugly. They were plausible. A cooler with the right silhouette but a slightly wrong latch. A logo that was readable but two degrees off. A six-second loop where the lid popped open for one frame. Those are the assets that eat a production day because they look ready enough to argue about.

Licensing caveats also added drag. Some tools allowed commercial use on lower tiers but reserved video ad rights for enterprise plans. Others required opt-out settings for training data, or limited likeness use if a model generated human faces. None of that showed up in the credit counter. It showed up in review.

The 'Almost' Tax

We started calling the gap between demo output and publishable output the almost tax. It includes prompt rewrites, inpainting, upscaling, frame cleanup, color matching, and the human time spent deciding whether an asset is close enough to save. Across the eight runs, edit time accounted for 61 percent of total CPUA on average. Tool spend was the visible line item; labor was the budget.

That matches what we hear from US creators and small product teams. A $30 monthly plan looks trivial next to a freelancer's day rate. But if the tool produces two usable assets per hour and your editor charges $75 per hour, the effective cost is not $30. It is the subscription plus the edit bay.

What We Would Do Differently

Next time we would lock acceptance criteria before generation, not after. We would log prompts and seeds in the same sheet as spend. We would separate prompt time from edit time so we can see which tool creates decisions versus chores. And we would run a 10-asset pilot before buying annual seats.

The headline from this audit is not that generative media is expensive. It is that advertised pricing hides the only number that matters for production: the cost of a keeper. Count the retries. Count the minutes. Count the week your legal reviewer spent reading terms. Then divide by assets you would actually put in front of a customer.