--- title: "Vercel Ship Put Agents on Stage and Sold a Token Card With No Public Price" show: "The Beat Studio" episode: 8 published: 2026-10-16 duration: "15:51" audio: https://thebeat.dev/podcast/2026-W42-vercel-ship.mp3 canonical: https://thebeat.dev/podcast/2026-W42-vercel-ship/ --- # Vercel Ship Put Agents on Stage and Sold a Token Card With No Public Price Vercel closed its Ship 2026 tour with Token Pass, thirty dollars a month for eighty dollars of open-model tokens spent at a third-party rate that is not on the public price list, beside a same-day report saying the money is moving to open models. Will and Brian then go through what the stage shipped (Python GA, Daemons, Drives, BYOC on AWS), the deployment-deletion fine print from the week before, and argue whether Vercel is now an agent cloud or a gateway testing what it can charge. ## Sources - [Vercel changelog - Introducing Token Pass](https://vercel.com/changelog/introducing-token-pass) - [Vercel docs - Token Pass](https://vercel.com/docs/ai-gateway/token-pass) - [Vercel docs - Token Pass credits and billing](https://vercel.com/docs/ai-gateway/token-pass/billing) - [Vercel docs - Token Pass models and routing](https://vercel.com/docs/ai-gateway/token-pass/models-and-routing) - [Vercel docs - AI Gateway pricing](https://vercel.com/docs/ai-gateway/pricing) - [Vercel AI Gateway - public models API](https://ai-gateway.vercel.sh/v1/models) - [Vercel - pricing](https://vercel.com/pricing) - [Vercel - AI Gateway Production Index, October 2026](https://vercel.com/blog/ai-gateway-production-index-october-2026) - [Vercel - AI Gateway Production Index, September 2026](https://vercel.com/blog/ai-gateway-production-index-september-2026) - [Z.ai docs - GLM Coding Plan overview](https://docs.z.ai/devpack/overview) - [Vercel Community - Registration is live for our biggest Ship yet](https://community.vercel.com/t/registration-is-live-for-our-biggest-ship-yet/39797) - [Vercel - Vercel Ship 2026 recap](https://vercel.com/blog/vercel-ship-2026-recap) - [Vercel - Python on Vercel is now generally available](https://vercel.com/blog/python-on-vercel-is-now-generally-available) - [Vercel - Backends on Vercel, Python, containers, and Daemons](https://vercel.com/blog/backends-on-vercel-python-containers-and-daemons) - [Vercel changelog - Vercel Daemons is now in public beta](https://vercel.com/changelog/vercel-daemons-is-now-in-public-beta) - [Vercel docs - Daemons](https://vercel.com/docs/daemons) - [Vercel - Introducing Vercel Drives](https://vercel.com/blog/introducing-vercel-drives-persistent-storage-for-agents) - [Vercel changelog - BYOC on AWS is now generally available](https://vercel.com/changelog/bring-your-own-cloud-on-aws-is-now-generally-available) - [Vercel changelog - MCP Events in Vercel Connect](https://vercel.com/changelog/mcp-events-support-in-vercel-connect-is-now-in-public-beta) - [Vercel - Vercel for Enterprise](https://vercel.com/blog/vercel-for-enterprise) - [Vercel - home page](https://vercel.com/) - [Vercel - Vercel Ship 2025 recap](https://vercel.com/blog/vercel-ship-2025-recap) - [GIC newsroom - Vercel Closes Series F at $9.3B Valuation to Scale the AI Cloud](https://www.gic.com.sg/newsroom/all/vercel-closes-series-f-at-9-3b-valuation-to-scale-the-ai-cloud/) - [Fly.io - home page](https://fly.io/) - [Railway - home page](https://railway.com/) - [Render - home page](https://render.com/) - [Vercel changelog - Deployment Storage billing begins for existing Pro teams](https://vercel.com/changelog/deployment-storage-pricing-expands-to-existing-teams) - [Latent Space - OpenRouter, from Seed to Stripe](https://www.latent.space/p/openrouter) - [The Pragmatic Engineer - Building Pi, and what makes self-modifying software so fascinating](https://newsletter.pragmaticengineer.com/p/building-pi-and-what-makes-self-modifying) - [Latent Space - Giving Agents Computers, Ivan Burazin, Daytona](https://www.latent.space/p/daytona) - [Code to Market - Secretive Marketing vs Loud Marketing?](https://codetomarket.fm/episodes/secretive-marketing-vs-loud-marketing) - [Scaling DevTools - Great DevRel Content is a process, not a project, with Jason Lengstorf](https://podcast.scalingdevtools.com/episodes/jason-lengstorf) ## Transcript Times are approximate (each turn placed by its share of the spoken words). **Brian** [0:06]: Thirty dollars a month buys you eighty dollars of tokens, and the public price list doesn't show what a token costs. That's Token Pass, and Vercel put it on sale this week. Vercel, if you've only ever met it through Next.js, is the hosting company behind that framework. It also runs a gateway for AI models: one key, and your agent can call any of the model companies through it, at that company's own price, with Vercel adding nothing on top. That was the whole promise. And this week that gateway started selling tokens at a price Vercel sets. If you've got a coding agent running all afternoon, this is aimed at you. Three things, then. The price list, which is mine. What the keynote actually put on stage, which is Will's, because he believes it. And a bit of fine print from the week before that deletes things. **Will** [0:58]: I believe some of it. Go on. What am I actually buying? **Brian** [1:02]: Here's the picture, and I'll keep coming back to it. Every month Vercel hands your team a prepaid card. It's loaded with eighty dollars. You pay thirty for it. The card works in exactly one shop, and that shop sells Kimi and GLM, both open-weight models, and nothing else. First odd thing about the shop: nothing on the shelves has a public price tag. The docs say the credit is spent at the provider's per-token prices, which they say are on the model page, and the provider isn't Vercel. It's a company called Wafer, which the docs describe, in so many words, as a non-Vercel service. The gateway publishes a rate for every model and every provider it routes to. Wafer's ordinary rates are on that list. What Wafer charges a Token Pass card is not on those public pages. So you walk in with the card, you pick things up, and the one thing the public pages won't tell you is what each of them cost. Now the terms. When the card's empty, your agent's requests fail until next month. No overage, no top-up, no quiet fallback to your normal paid key. What you don't spend doesn't roll over. It's for Pro teams only: Hobby has to upgrade first, and Enterprise can't have it at all. It's early access, a limited number of spots, with a waitlist once they're full. And Pro itself is twenty dollars a month. So the add-on costs more than the plan it hangs off. **Will** [2:30]: One shop, no price tags, and when the card's empty you're standing at the till with your arms full. Fine. But look at the front of the card, which is the only number they've given you, and it's more than double what you paid. Even blind, that's a discount, surely? Nobody prints a card that's worth less than what's on it. And think about who buys this. Not someone comparing rates. Someone who got tired of watching a meter run while an agent reads their codebase yet again. They don't want the rate. They want to know the card lasts the month. Does it last the month? **Brian** [3:07]: Depends what your agent does. Wafer's ordinary public price for Kimi is unusual: far below the going rate to read, and nearly twice it to write. So an agent that reads a lot of code and writes a little gets a long way on the card. One that writes a lot doesn't. And none of that is the Token Pass rate anyway, because that one isn't on the public list. Anybody who tells you the size of the discount is guessing. Which matters here because of what Vercel's own gateway pricing docs promise. The gateway, they say, "charges no markup and no platform fee on tokens". You pay the provider's list price, and Vercel earns around the token: zero data retention for the team, allowlists, reporting. Token Pass is the one thing on it where Vercel names the price and the rate underneath isn't on the public list. And the terms give Vercel room. It can swap the models and the providers, end the whole thing at any time, for any reason, with a pro-rata refund. And it can cancel your Token Pass without a refund for misuse, including, the terms say, through multiple accounts. **Will** [4:15]: Multiple accounts. Why would anyone... oh. Because if the card's worth more than you paid for it, you buy a stack of cards and sell what's on them. **Brian** [4:25]: That's the fear, and it isn't theoretical. OpenRouter, on Latent Space a few weeks ago, said anyone selling generalised inference is a target for fraud: people reselling traffic against the terms, accounts quietly reselling inference on the side, and the fraud they block had grown tenfold in a month. Their numbers, about their gateway. A card worth more than its price is exactly the thing you'd resell. **Will** [4:49]: Who's it for, then? Not the reseller, presumably. **Brian** [4:52]: The changelog names the coding agents it's built for, and the setup guide starts with ones you've heard of, Claude Code among them. No image models, no audio, no fast variants. And then the line that tells you the most: "For production traffic that has to stay up, use a regular AI Gateway API key." So this isn't for your product. It's for the developer in a terminal with an agent that burns tokens all afternoon. Mario Zechner, who built Pi, one of the agents the changelog names, said on The Pragmatic Engineer in the spring that what made Claude Code so enticing was the subscription, a flat price, and that paying by the token works in an enterprise but not for the small tinkerer in the garage. Token Pass is Vercel's offer to the tinkerer. A flat price, on open models, with the meter hidden. **Will** [5:43]: The tinkerer can get that from the model makers, though. Z.ai, who make GLM, sell their own flat coding plan for it, straight from the source, from about eighteen dollars a month, and the limits are on the page. So what's Vercel adding, other than Kimi? **Brian** [5:59]: Not a lot, on the face of it. The maker sells you the flat plan and tells you where the limits are. Vercel sells you the card and doesn't show you the rate publicly. What it adds is the gateway around it: a key inside a team that's already on Vercel. A convenience, or a habit, depending on who you ask. **Will** [6:20]: So it's a loss leader, aimed at people who'd otherwise be paying Anthropic? **Brian** [6:25]: Maybe. Here's where it gets interesting. The same day, Vercel published its AI Gateway Production Index, a monthly report on the traffic through its own gateway. The headline: open-weight models now take about one dollar in four of the spend, up from around one in twenty at the start of the summer. And Anthropic's share of the dollars, on their gateway, fell below half for the first time. Their line, in the report, is that the money is now following. Vercel's customers, Vercel's numbers, and they've changed how they count open-weight since last time. One source, and it's the vendor. Now look at which open models the money followed. Kimi led the growth in open-weight spend, the report says, and the pricier open models made up most of the rise. One of the two models on the Token Pass card, Kimi, is the one the same-day report says the money is moving to. The report never mentions Token Pass, and that overlap is my reading, not theirs. What makes me doubt it: the index comes out mid-month, every month. So the date may be the calendar, not the keynote. **Will** [7:31]: Then let me tell you what the keynote put on the table, because I think it was a real move, whatever the calendar did. Ship was a tour this year, and San Francisco was the last stop, where most of the previews became products. Here's the recap's first line, and it's the whole repositioning in one breath: "For a decade, Vercel has shaped how the web is built. Now, we're doing the same for agents." Agents first, apps second, and a backend to run them on. Three things came off that stage that I'd point at. Python is now generally available, FastAPI, Django, the lot, which is the plainest possible way of saying your backend can live here. Daemons: long-running workers that pull jobs off a queue, can't take inbound traffic, and on the free plan the allowance counts the CPU they use, not the time they spend waiting on that queue. An allowance that doesn't count idling is the right shape for a queue worker. And Drives: a disk that mounts into an agent's sandbox and outlives it, so an agent that moves environments keeps its files, and you can fork the disk for parallel workers. For the big accounts, your functions can now run inside your own AWS account, enterprise only, through sales. Put that together and it's a backend. It's what Render and Railway and Fly sell. **Brian** [8:51]: Does anyone actually run a backend there, or is this a brochure? **Will** [8:55]: A hundred and fifty thousand teams running Python backends on it, Vercel's number, nobody else's. Which tells you the label is catching up with the use. And the enterprise post is where the growth sentence is: enterprise token usage through the gateway quadrupled in a quarter, their numbers again, and the whole pitch is a control plane for every builder, app, agent, token and dollar. Then the home page, which has read agentic infrastructure since before the keynote, and puts agents ahead of apps. Gonto, on Code to Market recently, said that for devtools the UI is dying, and his co-host said a lot of the users don't have eyes. Vercel's front door is written for the user with no eyes. Look at the neighbours. Fly leads with computers for agents. Railway's home page is written as a briefing addressed to the agent reading it. Render leads with builders and gets to agents further down. Nearly everybody's headline says agents now. That's the stage. **Brian** [9:53]: And the week before all this, the same company published a changelog for its existing Pro teams with no keynote attached. It's the one thing in this whole week a listener might have to do something about. Storage for deployments is going to be billed, and the start date comes by email. And deployments older than thirty days get deleted from the twenty-third of October unless you've opted out in your retention settings before then. So say you've got a side project on Pro with an old deploy you were keeping as a rollback. The live deploy and the most recent ones are protected, the docs say. Anything older than that, past thirty days, is gone after that date, and the rollback button doesn't bring it back. **Will** [10:37]: It deletes the rollback? So the week before it told everyone it runs agents, it took your old deploys to the dump. **Brian** [10:45]: The ones you didn't say you wanted. A reasonable default for a new team, a quiet change for an old one. And it tells you where the attention is: on the agent that deploys every hour, not on the team that last deployed in the summer. **Will** [11:01]: Agent cloud, and here's why it isn't just a rename. Jason Lengstorf, who ran developer experience at Netlify, has a phrase on Scaling DevTools, borrowed from a restaurateur: constant gentle pressure. Being consistent with a pretty good message beats being inconsistent with great ones. Look at what Vercel did. Not one keynote. A tour, the same message in every city, previews in the spring, products in the autumn, and the same words on the home page since before the last stop. That's pressure, not a pivot. And it shipped the thing a rename can't fake. A disk that follows an agent between sandboxes is a real primitive; you can't slogan your way to one. So where's the money? Not in the card. In the enterprise post, where the growth is, and where Vercel charges for everything around the token. Token Pass isn't the business. It's the front door. Say you're the developer in the terminal, you've got Claude Code open, and you've never deployed anything on Vercel. You buy the card for the open models. Now you've got a Vercel team, a Vercel key, and a habit. A launch mostly reaches the people who'd have read the changelog anyway. The card is the one thing on that stage aimed at someone who wouldn't. **Brian** [12:15]: Your developer in the terminal. Walk him to the checkout with me. He can't buy the card. It's Pro only, and Hobby has to upgrade first. So he pays for a plan he didn't want, to buy a card he did, and the card is sold only to people already paying Vercel. That's not a front door. It's an upsell for the people already inside. **Will** [12:38]: So the card's the doormat, not the door? **Brian** [12:41]: The doormat. And real is doing a lot of work on your stage. Daemons are a public beta: production only, no autoscaling yet. Connect's new events are a beta. Containers were already in beta and Drives had one. Most of what came off that stage is betas graduating, and the worker you liked best hasn't graduated yet. You said it yourself: it's what Render and Railway and Fly sell. As for the message, Vercel's been calling itself the AI cloud for over a year, and raised its Series F on the phrase. By the spring swyx was already listing Vercel among the ones who say AI cloud. So I hear a GA day under a slogan that was already on the door. The thing that moved me, by the way, is that line steering production traffic off the card. If this were a business, they'd want the production traffic on it. They don't. So it's not a business. It's a probe. Can a gateway that promised never to touch the price charge for tokens, on open models, the kind its own report says people are moving to, without anybody able to check the rate? And they're running it on Pro teams, with a cap on the spots, and a clause that lets them stop. **Will** [13:56]: A probe is what you call it when you're waiting for proof, and who's going to hand you any? The public pages don't show the Token Pass rate. So I'll make the call now, with a date on it, which is more than a probe gets. The report and the card landing the same day, I take as planned, and I don't need it to be. A report that says the money's moving to the open models, next to a card that only buys the open models, is a marketing department with its stories straight, whether or not the calendar did the work. So, by next year's Ship: Token Pass is open to every Pro team, no cap, more models on the card, and the teams that bought it are deploying on Vercel. If that's what it looks like, then the card was the door and Pro was the doormat. If it's gone, or still a waitlist, you were right and it was a coupon. And the one thing I'd want from Vercel either way is the number on the shelf. Put the rate on the card. If the discount's real, say what it is. It sells better. **Brian** [15:05]: If they put the rate on the card, the card's just the gateway with a coupon, and the report's the brochure for it. I'll take the other side of your date. Until then, what Vercel sold this week is a prepaid card for a shop with no public prices on the shelves, and the first person to learn what the tokens cost will probably be whoever's agent stops at the till. --- Will and Brian are fictional hosts voiced 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. Part of [The Beat Studio](https://thebeat.dev/podcast/). Content license: CC BY 4.0 - quote it, link the canonical page.