What actually happened
Across the summer of 2026, the generative video category quietly finished converting from flat subscriptions to metered billing. The tools did not get more expensive on the sticker. They got harder to predict on the invoice.
HeyGen scrapped its unlimited plans on May 15 and rebranded its credits as Premium Credits, with the heavier avatar tiers (Avatar IV and V) burning about 20 credits per minute and enterprise contracts billed per generated second at roughly $0.50 per credit (eesel, Arcade). Runway prices by the generated second in credits, with Gen-4.5 at 25 credits per second (Runway pricing breakdown). Google's Veo, on Vertex AI, lists $0.50 per second for video and $0.75 per second with audio (CloudZero). Mid-tier text-to-video models run 20 to 60 cents per generated second at the API level, premium models 60 cents to several dollars (LTX).
None of this is a vendor turning greedy. It is the underlying economics surfacing. Generating video burns real GPU time, and for two years that cost sat hidden behind venture-subsidized flat plans built to win market share. Flat pricing works while a category is buying users. It stops working the moment the same vendors have to show a path to margin, and metering is how every compute-heavy category eventually turns usage into the customer's problem instead of the balance sheet's.
Then the second leg. The subsidy is ending. Promotional free-credit programs issued from 2023 through 2025 are expiring at scale, and analysts estimate the average brand's AI content bill rising 35 to 60 percent this year alone (Rewarx). Put together, the message is simple. The list price of AI video is no longer the price you pay. The meter is.
Why a cheaper per-second rate is the worse budget line
A metered tool can be cheaper per second and still be the harder thing to budget, because the cost you actually care about is not per generation. It is per usable asset.
Every botched generation bills at full price. There is no discount for the four takes you throw away to reach the one you can ship, and regenerating after the model misreads your prompt costs exactly what the first attempt did. A shippable 20 second clip is rarely one generation. It is the handful you discard, plus the human who still has to do the caption, the brand pass, and the accuracy check afterward. The per-second sticker rate describes the cheapest possible world, the one where the first generation is perfect. That world does not exist in B2B, where the brand bar is high and the subject is a real product a buyer will scrutinize.
Watch how that compounds through a normal review cycle. Marketing drafts a 30 second explainer, generates it, and sends it to product for an accuracy check. Product flags a wrong interface frame, so you regenerate. Legal wants a claim softened, so you regenerate. The founder wants a different opening line, so you regenerate. Every one of those rounds is a full-price generation, and none of them is waste in the ordinary sense. That is simply how B2B content gets approved. A flat retainer prices those rounds in. A meter charges you for each lap around the same track.
Now put that against a plan. A retainer gives you one number you can drop into a budget and defend to a CFO. A meter gives you a number that moves with prompt luck, model-version changes, and credit-rule updates you do not control. For a marketing leader, predictability is not a nice-to-have. It is the thing that lets you commit to a content calendar in the first place. You cannot build a publishing schedule on a line item that swings with the weather.
The data
Here is what the gap between sticker price and real cost looks like once you measure it.
In our retainer client audit, across more than 20 SaaS engagements, teams that shifted product cutdowns and social clips to credit-metered AI tools in late 2025 landed a true cost per usable asset around 2.4 times the sticker credit rate, once we counted regenerations and the human cleanup pass. Our production time-study on avatar and text-to-video clips found a median of four to seven generations to reach one usable 20 to 30 second asset, and every failed generation billed at full rate. In one case from our retainer network, a client's monthly AI video spend swung from about $180 to $1,400 across a single quarter with no increase in output, driven by regenerations and a mid-quarter credit-rule change on their plan.
The public benchmarks point the same direction. Independent pricing teardowns put premium generation at 60 cents to several dollars per finished second (LTX), and the expiry of subsidized credits is pushing the average brand's AI content bill up 35 to 60 percent this year (Rewarx). The sticker second is the floor, not the bill.
The counter-argument, steelmanned
The strongest case against this is that metering is honest and control is possible, and that is partly right. Metered pricing means you only pay for what you use. A disciplined team can cap spend, cache the outputs that land, and cut regeneration waste with better prompting and reusable templates. For high-volume, low-stakes assets like localized variations and quick social tests, metered AI is genuinely cheaper than any human workflow, and a retainer you underuse is its own kind of waste.
That holds for a specific asset class and a specific level of operational maturity. It breaks where B2B actually lives: brand-sensitive, product-accurate, founder-on-camera work where the acceptance bar is high and regeneration counts climb fast. And notice what "you can control it with discipline" is really saying. Discipline is labor. The meter does not remove the human. It relocates the cost from the invoice to your team's calendar, where it is harder to see and easier to underprice. The real retainer question was never flat versus metered. It is who absorbs the variance. A content system absorbs it for you. A meter hands it to you and calls it savings.
What to do Monday
You do not need to rip out your AI tools. You need to price them honestly and put a ceiling on the variance.
Start by measuring cost per usable asset, not cost per generation. Pull last quarter's credit spend, divide it by the number of assets you actually shipped, and compare that to the sticker rate. The multiple you find is your real number, and it is the one to put in next quarter's plan.
Then put a hard monthly ceiling on metered spend and set an alert at 70 percent. Treat a runaway regeneration loop like any other cloud cost that can quietly triple, because that is exactly what it is.
Sort your content by stakes. Route low-stakes, high-volume variations to metered AI, where it genuinely wins. Route brand-sensitive, product-accurate, on-camera work to a fixed-cost system where the price does not move when the model version does. Finally, when you renegotiate, negotiate for predictability, not just rate. Whether with a vendor or a production partner, the term that protects your plan is a fixed monthly envelope, not a slightly lower per-second price. The goal is not to avoid AI. It is to stop letting a meter you do not control set your marketing budget.