Dora NodaOpen a Vercel invoice from 2023 next to one from this year and almost nothing matches. Bandwidth is now Fast Data Transfer. Serverless execution split into Active CPU and Provisioned Memory.
The Pro plan is a seat fee stapled to a usage credit. Four pricing revisions in three years, each announced as simplification — and yet teams forecasting next year's bill keep arriving at the same uncomfortable totals. What changed was the vocabulary. What didn't is the economics.
Here's the translation table, up front, so you can stop decoding and start forecasting.
| What your 2023 bill said | What the 2026 bill says | Hobby included | Pro included | Did the price move? |
|---|---|---|---|---|
| Bandwidth | Fast Data Transfer (+ Edge Requests) | 100 GB | 1 TB, then overages | No — same caps, same overage class |
| Serverless execution (GB-hours) | Active CPU + Provisioned Memory | 4 hrs CPU + 360 GB-hrs mem | Metered from $0.128/hr CPU | Yes — the one genuine cut |
| Pro subscription ($20/seat) | $20/seat + $20 usage credit | n/a (free, non-commercial) | Credit nets against usage | Restructured, not reduced |
| Function invocations | Function invocations | 1 M | Metered (~$0.60/M where billed) | No |
And here's what that means in dollars for a typical five-person team at three traffic levels, versus the same workload on a flat-price box. Assumptions are spelled out below the table; the point lands faster than the footnotes.
| Monthly transfer | Vercel Pro (5 seats, base region) | Flat Hetzner box (AX41-class) | Delta |
|---|---|---|---|
| 1 TB | ~$170 | ~$55 | ~$115 |
| 2 TB | ~$320 | ~$55 | ~$265 |
| 5 TB | ~$770 | ~$55 | ~$715 |
The flat line doesn't move because there is no meter to rename. Everything below is the receipt for how we got here — revision by revision — and what to watch when you forecast.
Revision 1 (2024): "Bandwidth" becomes Fast Data Transfer
The first rename landed in 2024, when Vercel broke the single bandwidth line into granular meters: Fast Data Transfer for bytes served, Edge Requests for request count, plus cache read/write units for ISR and image content. Hobby kept 100 GB of transfer; Pro kept 1 TB. The caps didn't move. The per-unit economics didn't move either — Pro overages stayed in the $40-per-100GB class that had already made bandwidth the most complained-about line on the invoice.
What the split clarified is real: cache hits and origin fetches are different costs, and itemizing them lets a team see whether its bill comes from serving bytes or from revalidating them. A site with a 95% cache-hit ratio and a site doing per-request SSR finally get different-looking bills, which is genuinely useful for optimization.
What it obscured is that the optimization target didn't get cheaper. Whether the line says "bandwidth" or "Fast Data Transfer plus Edge Requests plus cache units," a team pushing 2 TB a month pays overage rates per gigabyte that are two orders of magnitude above raw transit.
Regional per-GB pricing (roughly $0.15 to $0.40 per GB depending on region) means the exact multiple varies, but the shape is the same: transfer is the meter that turns a traffic spike into a bill spike. Granularity helps you explain the bill. It doesn't shrink it.
Revision 2 (2025): GB-hours split into Active CPU and Provisioned Memory
The second revision is the one genuine price cut in the whole sequence, and credit where due. With Fluid Compute going mainstream and Active CPU pricing announced at Ship 2025 (enabled by default that June), Vercel stopped charging compute rates for time your function spends waiting on I/O. The old GB-hour meter billed provisioned capacity for the whole invocation lifetime; the new model bills Active CPU (from $0.128/hour) only while code executes, plus Provisioned Memory (from $0.0106/GB-hour) for the instance's lifetime. Vercel claimed up to 85% savings for streaming and AI-inference workloads — the exact workloads where functions idle on network waits.
For Hobby, the translation is concrete: the old 100 GB-hour serverless allowance became 4 hours of Active CPU plus 360 GB-hours of Provisioned Memory. Those numbers look small next to "100 GB-hours" until you realize CPU-active time is a fraction of wall-clock time for I/O-bound functions. A chatbot streaming tokens for 30 wall-clock seconds might burn 3 CPU-active seconds. The new meters fit that shape; the old one punished it.
The catch is scope. Active CPU cut compute, which for most frontend-plus-API teams was never the scary line. Nobody's $400 surprise bill was 90% function duration.
The scary lines — transfer, seats, invocations at scale — kept their rates. So revision 2 is simultaneously the most technically admirable change (pay for execution, not idle) and nearly irrelevant to the bill-forecast question most teams actually ask: "what happens when we get featured on Hacker News?" The answer is still: the transfer meter spins.
Revision 3 (September 2025): Pro becomes a seat plus a credit
In September 2025 Vercel restructured Pro around a flexible spending model: still $20 per deploying seat per month, but each team month now starts with $20 in flexible credit usable across products, plus dedicated allowances (about 1 TB of Fast Data Transfer and 10M edge requests at the base tier), with free read-only viewer seats for everyone who doesn't deploy. Spend Management alerts turned on by default, with a $200 on-demand limit for new teams.
What this clarified: one balance instead of 20+ discrete allotments to track, and an end to paying $20 for a seat whose owner only reads preview links. Vercel said over 100,000 teams would see bills decrease or hold flat, with around 7% seeing increases.
What it obscured: the $20 everyone quotes is now doing double duty as both seat fee and credit funding, and seats still multiply without volume discounts. Five developers is $100 before a single visitor arrives; ten is $200. The $20 credit nets against usage overages, which softens small overruns but is a rounding error next to a real transfer spike — one extra terabyte at base rates ($150) eats the credit seven times over. And "Enterprise features available self-serve" (SAML SSO, HIPAA BAAs as paid add-ons) quietly moved the ceiling: line items that used to require a sales conversation now just appear on the card.
Revision 4 (2026): Hobby settles — with a rule that ends side projects
By 2026 the Hobby free tier stabilized at its current shape: 1M edge requests, 100 GB Fast Data Transfer, 1M function invocations, 4 hours Active CPU, 360 GB-hours Provisioned Memory, 1 GB Blob storage, one member, unlimited deployments. No build minutes, no purchasable overages — hit a cap and the allowance stops rather than billing you, which at least makes Hobby predictable.
But Hobby carries the clause that catches more people than any rate: personal, non-commercial use only. The moment a side project earns revenue, it's out of compliance with the free tier, and the hop to Pro isn't "$20" — it's $20 per deploying seat plus metered everything.
For a solo founder that's a manageable step. For the three-friends side project that just got its first paying customer, it's $60 in seats before usage, on a project that made $19 this month. Every pricing page buries this; every forum thread about surprise Vercel bills eventually finds it. When you forecast, price the compliant tier for the project you hope to have, not the free tier for the project you have today.
What stayed put: the three lines that decide your bill
Strip the four revisions down and three economics never moved:
- Transfer overages. Pro includes 1 TB; beyond that you pay per-GB rates in the $40-per-100GB class (regional pricing runs roughly $0.15–$0.40/GB). A media-heavy page, an unoptimized image pipeline, or one viral day converts directly into hundreds of dollars. This is the meter behind nearly every public Vercel-bill horror story, including the March 2026 migration writeup where a 2 TB month produced over $330 in transfer lines alone.
- Seats multiply. $20 per deploying developer, no volume curve, viewer seats free. Your headcount sets a floor under the invoice that traffic can't take away — the inverse of usage-based pricing's promise.
- Everything else is metered. Edge requests ($2/M past 10M), invocations (~$0.60/M where billed), builds per minute, storage, firewall rules past the allowance. Individually small; collectively the reason a "simple" bill has a dozen lines.
Here's the worked example behind the opening table, so you can audit it. Five-person team, 2 TB transfer, 50M edge requests, moderate function usage (~40 Active CPU hours, ~500 GB-hours memory), base region:
- Seats: 5 × $20 = $100
- Fast Data Transfer: 1 TB over × $0.15/GB = $150
- Edge requests: 40M over × $2/M = $80
- Compute: 40 × $0.128 + 500 × $0.0106 ≈ $10
- Less $20 flexible credit = −$20
- Total: ~$320/month
At 1 TB the transfer line drops to zero ($170 total). At 5 TB it quadruples ($770). And in a $0.40/GB region, multiply every transfer line by ~2.7x — the 2 TB case becomes roughly $570. That sensitivity to region is the part no rename can fix: it's the underlying unit price.
The self-hosted counterfactual: the same workload on a flat box
Now the same workload on owned-style infrastructure: an AX41-class Hetzner dedicated box (Ryzen 6-core, 64 GB RAM, NVMe, 1 Gbit uplink) at roughly €40–60/month — call it ~$55. Flat. No transfer meter (bandwidth is included, not itemized), no seat count, no per-invocation line.
1 TB, 2 TB, 5 TB: the invoice reads the same. That's the flat row in the opening table, and the delta column is the price of the meters.
This isn't hypothetical. The Rootz migration writeup from March 2026 walks through a real $400/month Vercel bill — $200 in Fast Data Transfer, $136 in origin transfer, single-digit compute — moved to a ~$40 VPS. The pattern matches the math above exactly: compute was never the problem; transfer was the whole bill, and transfer is the line a flat box deletes.
Honest caveats, because the delta isn't free. The flat box doesn't come with a global edge network, zero-config previews, or someone else's on-call rotation. You trade the meters for ops time: OS patches, deploys, TLS, scaling past one machine.
For a team with no ops appetite, Vercel's premium buys real convenience, and preview deployments alone are worth money. But notice what that reframes: you're no longer comparing "hosting vs hosting." You're comparing a hosting bill against an ops-time budget — and for a standard frontend-plus-API workload, the hosting half of Vercel's bill is mostly transfer margin.
There's a middle path that keeps the git-push workflow while deleting the meters: a self-hosted PaaS on that same flat box. Push-to-deploy, preview URLs, managed TLS — without a per-seat fee or a per-gigabyte transfer line, because the platform runs on machines you already pay a flat rate for. That's the architecture worth pricing against Vercel, not a hand-rolled VPS with a README deploy process.
Forecast the meters, not the marketing
If you're budgeting Vercel for next year, ignore the revision announcements and watch four numbers: transfer volume against the 1 TB Pro inclusion (and your regions' per-GB rate), edge-request count against 10M, deploying-seat count times $20, and whether your Hobby projects are still non-commercial. Those four lines decide 90% of real invoices. Everything else — Active CPU, provisioned memory, cache units — is a more precise way of measuring the cheap parts.
Four revisions made the bill legible. Only one made it smaller, and only for compute. The next time a pricing page announces new meters, ask the question this whole sequence answers: did the unit price move, or just the unit?
Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own, with no per-seat fee and no transfer meter. Star the repo on GitHub or deploy your first app today.
