My May Vercel invoice landed at $312.40 and I stopped pretending I needed it.

This is a small B2B SaaS. Maybe 4,000 monthly active users, a marketing site, a docs portal, three serverless cron jobs. Nothing exotic. The kind of app any Hetzner box would yawn through. So why was I paying enterprise prices for a Next.js skin?

I gave myself a Saturday. Six hours. No new tools I had to learn from scratch — only Cloudflare Workers, Supabase, R2, and Resend, all of which I’d touched before. The brief I wrote on a Post-it stuck to my monitor read: zero invoice by Monday morning, or admit Vercel was worth it.

What I was actually paying for

Vercel Pro is $20/seat/month. That bit’s fine. The other $292 was the part that quietly bled me: a credit-based billing system Vercel rolled out in September 2025 that turned bandwidth, function invocations, and image optimization into separate line items, each with its own meter and its own surprise.

Then in February they made Turbo build machines the default for new Pro projects. Turbo is 9x the price of Standard. They emailed nobody. I noticed because my deploy line item went from $11 to $94 in one month.

Brutal.

Translation: the predictable ceiling is gone. You get $20 of “flex credit,” then everything’s a meter. Convenient for them. Less so for the spreadsheet I send my accountant.

This connects to our analysis of which SaaS categories won’t survive AI compression. Vercel’s premium tier is on that watch list — not because the product is bad, but because the marginal value of “managed Next.js” is collapsing as the open-source adapters catch up.

The four free tiers I bet on

I picked four services that I’d already integrated before, so the migration was about moving not learning. Here’s the swap:

Cloudflare’s daily request limit on free is 100,000. My app does about 38,000. Headroom: comfortable. Supabase free gives 500 MB of database storage on a shared compute box, which sounds tiny until you realize my entire production Postgres is 217 MB after three years. (More on why Supabase keeps winning the YC default slot in our Supabase vs Firebase breakdown.)

The actual six hours

1. 08:14Coffee. npm i -g wrangler. Cloudflare account already existed from a side project. ~12 min.
2. 08:26Adapted Next.js app to @opennextjs/cloudflare. One config tweak, one runtime export per route. ~58 min.
3. 09:24Postgres dump from Vercel via pg_dump. Restored to a fresh Supabase project. Same Postgres flavor, zero schema drift. ~41 min.
4. 10:05Auth.js → Supabase Auth. This was the only painful part. Two hours of session-cookie debugging because I forgot the JWT secret had to match the old issuer.
5. 12:18Migrated 4.2 GB of user uploads from Vercel Blob to R2 with a shell loop. Cloudflare egress to R2 was free.
6. 13:30Resend already in place. Just swapped the API key and pointed the SPF/DKIM at the same verified domain.
7. 14:11DNS flip. Cut over with a 60-second TTL window. Zero user-reported errors in Sentry.
8. 14:28Deleted the Vercel project. Slightly emotional moment. Closed laptop.

The OpenNext adapter for Cloudflare did most of the heavy lifting. Two years ago this would have been a fortnight of yak shaving. In 2026 it’s an evening.

The one config that broke me

If you’re following along, the gotcha was middleware. Vercel’s middleware runs at the edge. Workers run everything at the edge by default. So my old middleware-gated routes were getting double-handled until I deleted the file outright. The runtime declaration that finally made it work:

// next.config.ts
import { setupDevPlatform } from '@opennextjs/cloudflare/dev'
if (process.env.NODE_ENV === 'development') {
  await setupDevPlatform()
}

export default {
  experimental: {
    runtime: 'edge',
    serverComponentsExternalPackages: ['@supabase/supabase-js'],
  },
  images: { unoptimized: true }, // Workers handles this via R2 + cf-image-resizing
}

Note the images: { unoptimized: true }. Vercel’s image optimization is a known billing trap (one widely-cited case study shows a $20 plan ballooning to $286). Cloudflare’s image resizing is included with your Workers traffic at no marginal cost up to 5,000 unique transformations per month on the free tier. Plenty for me.

What the new bill actually looks like

The math, three weeks in: I haven’t paid a cent. I came within 22% of the Cloudflare daily request cap once during a HackerNews mention, and Cloudflare let me through anyway because their throttling is a 429 not a hard wall — it just degrades. A paid Workers plan kicks in at $5/mo if I cross 10M requests in a month, which I won’t.

The blind spots

Three things I didn’t test, in case you’re tempted:

• I didn’t load-test past 100K req/day. My traffic is steady. If you spike, you’ll need a paid Workers plan or a queue layer. The bill stays small but it isn’t zero.
• I didn’t migrate ISR pages. The OpenNext adapter supports it now but I rebuilt my few cached pages as Worker static routes. Cleaner. Your mileage will differ.
• I didn’t run this on Windows. I can’t be bothered. Wrangler on WSL works fine from what I’ve read, but I’m a Mac person, sorry.

The pricing structure now emphasizes flexibility with the credit system while adding more paid add-ons for advanced features that were previously enterprise-only.

Vendr’s 2026 Vercel pricing summary — read “flexibility” as “unpredictability”

Who should not do this

If you’re a 12-engineer team shipping 40 PR previews a day with shared environments and SSO, Vercel earns its keep. The preview infra alone is a quarter of an engineer’s time. Don’t migrate.

If you’re a solo dev paying $300/month because you set things up two years ago and never looked again — spend the Saturday. Worst case you go back. Vercel’s import tool reads a Cloudflare project and restores in 15 minutes. The exit is symmetric.

What happens next

Vercel knows. The September 2025 repricing was their first acknowledgement that Pro had drifted upmarket. The February Turbo-by-default move was the second. I’d bet on a $5 “Indie” tier landing by AI Engineer SF this June — capped resources, predictable bill, designed exactly for the migration I just did so they can keep the long tail from defecting. If it lands, I might even come back. The credit system, though? That has to die for me to mean it.

Until then, my Saturday saved me $3,749 a year. I’ll take it.