The Billing CEO Who Tried to Kill the Token and Kept It

3 October 2026 12:49 Will & Brian

A billing CEO whose company Stripe now owns tried to kill token billing and kept it, and says why. Will reads the post and grades it; Brian asks whose argument it is, and answers with GitHub's credits. Then DigitalOcean puts a subscription on the same meter, and Google's new Gemini model launches at a price with no end date.

Sources

Transcript

Click a time to jump there. Times are approximate.

The man whose company bills the token sellers tried to kill token billing, and kept it. That's the story this week, and if there's a model under your product it's yours too, because the question is which unit goes on your pricing page: the one you pay for, or the one your customer gets. Scott Woody runs Metronome, the usage-billing company Stripe bought at the start of the year, and he's just published a post on Stripe's blog titled why I tried to kill token billing and why we kept it. A billing vendor, in writing, saying the thing he meters is a bad way to charge a customer. Then shorter ones. DigitalOcean launched agent infrastructure on a per-second price and, a week later, put a monthly subscription over the same price. And Google announced a model at an introductory price that doubles when the introductory period ends, without saying when. So, Woody. Why would a billing company want to kill the thing it bills?
It's better than that sounds. He never wanted to stop metering tokens; he wanted them off the customer's invoice. The post says your invoice should be where you define your value to your customer, and that pricing on tokens means letting your costs determine your price. Here's the argument, with one picture I'll keep all episode. Imagine a restaurant that bills you for flour. Not the bread, the flour: so many grams, at the mill's price, plus the kitchen's markup, itemised. That's a token invoice. Woody's point is that once the customer can see the mill's price, they argue about the markup. And when flour gets cheaper, and models keep getting cheaper, they ask why their bill didn't, or they go and buy flour themselves. His sentence for it, the one I'd keep from the whole post: "Token billing positions your product as a commodity markup on top of an actual commodity." He's fair about why everyone does it. Tokens are a safeguard against runaway compute, and if you sell models the token is the product, so it's the right unit. For everyone else he calls it useful infrastructure and a bad thing to show the customer, which is why the title says he kept it: Metronome still meters tokens underneath, it just wants them off the bill. Then he gives you three steps. Tokens, which is the flour. Unified credits: a prepaid card, and each dish takes its own number of points off it, so the card knows what the flour cost and you don't have to. Then outputs: a price per dish, an image generated, an email sent. And outcomes, paying for revenue or churn, he calls a myth for anyone without a monopoly or a very large contract, because nobody can attribute them. My grade: as writing, high. As advice, a notch lower, and I think you're about to tell me why.
Not yet, hold the grade for a minute. So what does he want me to change on Monday?
Put the unit the customer values on the invoice, and keep the unit you pay for in the cost model. That's the principle, and I'd sign it. Whether a devtool can actually do it is where my grade drops, and I'll get there. You first, because I can see you've got the acquisition page open.
Mine's the listener's question too. Who is telling me this, and what does he sell? Stripe finished buying Metronome at the start of the year, and in the announcement Patrick Collison named the companies Metronome's metering runs for, OpenAI and Anthropic among them. The token sellers. The post ends by linking to Metronome's product page. So a billing vendor whose showcase customers bill in tokens is telling everyone else to buy a layer of billing on top of tokens. Doesn't make him wrong. So weigh it: it's vendor marketing. And here's what GitHub did in the summer. Copilot moved off a price per request, back toward tokens, wrapped in credits: every interaction priced by tokens and converted into credits, a credit a cent. Paid plans get a fixed batch of base credits plus what GitHub calls a flex allotment, and the docs say that flex batch is designed to adapt as the economics of AI evolve. The credit stays a cent; the docs say how many you get can move. That's the terms. My reading: a price that can change without the price changing. Woody's middle step, the prepaid card, is also a way to change how many points come on the card without ever touching the price of a point.
Change the points, not the price? That's a good line and I hate it. Is there a version of credits that can't do that, or is that just what a credit is?
PostHog's is the nearest. It runs its AI features on credits, and its docs print the markup: twenty percent over what the model provider charges them, written down on the page. The exact thing Woody says never to show, in public, on purpose. Then the one figure in the post. Woody says roughly one in six Stripe users past some revenue milestone are using or rolling out hybrid pricing. Hybrid, in his own paragraph, means seats plus usage. Not credits, not outputs. So the number counts a thing the post isn't arguing for, with no base given. Self-reported, single source, and it measures the wrong step, which tells you how thin the evidence for the middle one is: a billing company's own customers, counted in a way that leaves it out.
That takes my advice grade down another notch, and here's where it stops, because the answer is on the record and isn't from a billing company. Sarah Sachs, Notion's head of AI, on Latent Space in the spring. Notion's first idea was to charge per agent run, a price per dish. Every version of it, she says, brought them back to complexity that mapped straight onto token throughput. So they built credits above tokens, because their costs arrive in different units, GPUs, web search, serving tiers, and they sell credit packs with enterprise discounts. Self-reported, one company, but one that tried the step above and came back down. So here's what should have been done, on Notion's evidence, and it's one position: credits, tied to your cost in public, the way PostHog prints its markup. Then the number of points on the card can't move unless the formula does, and the formula is on the page where your customer can read it. For a devtool's pricing page that means the dish is what the customer sees, the flour stays in the kitchen, and the recipe for the card is published. That's the version of Woody's post I'd hand a founder, and it's shorter than his.
I had credits down as a way to hide a price. Notion tried the honest unit, per run, and it ran back to tokens on its own. That isn't a vendor hiding a price, it's a product that couldn't find a unit, and that is what moved me. But how do you know any of it works? Every position we've got is an opinion or a company describing itself. Has anyone measured whether tokens or credits move sign-ups, or bill shock, or anything a finance team would recognise? Not Woody, not Notion, not GitHub, nobody I could find. Which brings me to a company that had this whole argument with itself in one week.
DigitalOcean. Go on.
DigitalOcean, the cloud for small developer teams, launched Managed Agents into public preview with a post by its chief product and technology officer. The pitch was the meter: CPU billed per second of actual use, memory by the hour, and the post ends on being billed for the work and not the wait. The same post says infrastructure is roughly three to eight percent of what a typical run costs and the rest is tokens. Which tells you the per-second meter sits on the smallest slice of the bill. A week later, Agent Droplets. A monthly plan over the same meter: a Pro tier at fifty dollars a month with fifteen percent off agent usage, unlimited seats, no per-seat charge. The per-second rates did not change. Past the plan's allowance you pay list price from a prepaid balance. In your restaurant it's a membership card: pay monthly, get a discount on the menu, and the menu stays the menu. Their reason is predictability, not cost. They tell of one customer's stack, half a dozen vendors and as many invoices, in their own telling, and their picture of that month is labelled illustrative. Then: "Nobody on the team could say what a single agent run had cost." Small teams, they say, want a price you can predict and a bill you can read.
A monthly fee for a discount on something they already sell by the second. Hang on. Same rates underneath, a cheaper rate for paying up front. Why is that a story and not a coupon?
Because of the footnote. The per-second meter, billed for the work and not the wait, isn't live. The runtime's pricing page says billing only for the seconds the CPU is actually working is coming soon, and until then you're billed at a quarter of the CPUs you allocated, working or waiting. So in one week: a meter that launched, a membership that discounts the meter, and a line in small print saying the meter is still being installed. The membership arrived before the meter did.
Here's why I'm on DigitalOcean's side anyway, and my witness is a company going the exact opposite way. Depot, the build-acceleration tool. Kyle Galbraith, its co-founder, on Scaling DevTools last year: they were moving off their monthly tiers and back to pure per-minute usage, because a plan forces engineers to think about the price, with spend caps on top. Self-reported. He said they'd built a pricing model that worked for them as founders and didn't land with them as engineers. So the devtool took the membership card off, DigitalOcean put one on, and both say they're fixing the same complaint. Which tells me the unit isn't the argument. The argument is whether anyone can say in advance what the month costs. Hank Taylor, when he was at Laravel, on Code to Market: not knowing what it would cost was one of the main three reasons people hadn't made a first deployment, and his team was building a calculator. Self-reported, no figure. Predictability is the product. The unit is a detail you argue about afterwards.
Depot is one founder on one podcast, and you've just made him the industry. How do you know it's predictability and not margin? A plan with a discount is also a prepayment: DigitalOcean gets the monthly fee whether you run an agent that month or not. Do I need a predictability story to explain why a cloud company likes that shape? What I'll give you: Depot and DigitalOcean disagree on the unit and agree on the complaint, from both sides of the counter, and that's the most useful thing anyone has said today.
Last one, short, because you can't buy it. Google announced its new Gemini model, Argon, at an introductory price, and in small print the price doubles after the introductory period. No end date anywhere on the page. The model goes first to a cyber-defence programme, then to paid API customers, and it isn't on the public API price list yet. A sale price on flour, from a mill that isn't taking orders. Peter van Hardenberg, from Heroku, on To Be Continuous years ago, self-reported and old: "People really understand your product through its price." Heroku treated a price change as something the whole company had to roll out with care, and this one arrives as a line under a launch post. So what does a developer understand from a price that's half of something, for an unknown time? That the token price is a launch promotion. And this is my view, not Woody's: even the people who sell tokens use the token price as marketing, which is one more reason not to copy it onto your own invoice.
The flour's on sale and the mill isn't open.
So. Dish on the bill, flour in the kitchen.
And the recipe on the page, or someone quietly changes the portions.
Print it, and predictability is the product.
Then show me the measurement. From anyone.
Coming soon.

Will and Brian are fictional hosts voiced with ElevenLabs, and the music is made with ElevenLabs; the episode was researched and written by AI agents and fact-checked against the sources above. ElevenLabs is a company this site covers; it is a tool the site pays for, not a sponsor.

transcript as markdown mp3 all episodes