Dora NodaIn September 2025, Vercel quietly changed what counts as a "seat" on its Pro plan — teams now start with one included paid seat, get free Viewer-only seats, and get one free Billing-role seat. That wasn't an isolated tweak. It was the fourth time since early 2024 that Vercel rewired a core piece of how its bill is calculated, and each of the four landed on a different part of the invoice: bandwidth and function metering in 2024, the compute-execution model in February 2025, the CPU-billing unit in June 2025, and seats in September 2025.
Four repricings in about eighteen months is the fastest cadence among the major git-push platforms. Here's the full timeline, what each of the other three — Render, Railway, and Fly.io — did in the same window, and what a platform that keeps rewriting its own pricing model tells you about the cost pressure sitting underneath it.
The Vercel Timeline: Four Events, Four Different Line Items
| Date | What changed | Concrete effect |
|---|---|---|
| Apr–Jul 2024 | Bundled bandwidth/CPU allocations split into granular per-dimension metering — bandwidth, edge requests, function invocations, CPU hours, memory GB-hours, image optimization, and build minutes each become separate billed line items, with some per-unit rates cut | Bills fragment from a handful of pooled quotas into seven-plus tracked dimensions |
| Feb 12, 2025 | Fluid Compute launches: one function instance can now serve many concurrent requests, killing the old Edge Functions vs. Serverless Functions split and the per-invocation cold-start surcharge | Replaces the core billable unit — still time-based, but the unit of "compute" itself changes |
| Jun 12, 2025 | Active CPU pricing for Fluid Compute ships: you're billed only for CPU-active milliseconds, and I/O wait time costs $0 | Up to a 90% cut for I/O-heavy workloads — AI inference calls, database-bound requests — that spend most of their wall-clock time waiting, not computing |
| Sep 15, 2025 | Pro plan overhaul: 1 included paid seat by default, free Viewer seats, one free Billing-role seat, a $20/month flexible-spend credit that auto-absorbs overages (bandwidth → edge requests → function invocations, in that order), plus self-serve SSO/HIPAA BAAs that used to require Enterprise | Redefines who "counts" as a billable seat — a team with five reviewers and two deployers now pays for two seats, not seven |
The cadence hasn't slowed since. Into 2026, Vercel shifted function-invocation billing from a bundled $0.60-per-million-invocations package to raw per-unit pricing at $0.0000006 per invocation (the same effective rate, but a structurally different bill), and layered on new metered BETA SKUs for its Agent product (token-metered at $0.25 per million tokens plus pass-through model cost), Vercel Services, Queues, and Container Registry. None of that is a fifth "repricing" in the same sense as the four above — it's new usage-based products, not a rewrite of an existing bill — but it confirms the same platform is still actively adding metered surface area, not settling into a stable model.
Four times in eighteen months, a team on Vercel had to re-learn which line item was about to dominate its invoice. That's the pattern this post is actually about — not whether any single change was reasonable in isolation (several genuinely lowered costs for specific workloads), but how often the shape of the bill itself gets redrawn.
What Render, Railway, and Fly.io Did in the Same Window
Render: one confirmed major repricing, April 23, 2026. Render dropped its per-seat workspace plans — Professional at $19/seat/month, Organization at $29/seat/month — for flat monthly fees: Pro at $25/month flat, Scale at $499/month flat. The bigger line-item change was bandwidth: included egress dropped from a flat 100 GB across legacy plans to 5 GB on the new Hobby tier and 25 GB on Pro, with overages now billed per GB instead of in 100 GB blocks. Render says 75% of paying customers see costs decrease or stay flat under the new structure; legacy workspaces have until August 1, 2026 before they're auto-migrated. That's the only major, dated repricing event in Render's 2024–2026 window — one clean event, not the two the "every platform reprices constantly" narrative sometimes assumes.
Railway: zero repricing events. The Hobby ($5/month) and Pro ($20/month) base fees haven't moved since 2024, and the underlying usage rates — $0.000231 per GB-minute of RAM, $0.000463 per vCPU-minute — are unchanged too. Railway's last real pricing disruption was killing its free tier in mid-2023 after crypto-mining abuse, which sits before this window entirely. In a market where every other major platform touched its pricing model at least once, a flat sticker price for two-plus years is itself the data point.
Fly.io: three incremental additions, no full model rewrite. Fly.io removed its permanent free tier in 2024 (new orgs get $5 in trial credit instead), rolled out granular inter-region data-transfer rates for organizations created after July 18, 2024, and — starting January 1, 2026 — began billing for volume snapshot storage at $0.08 per GB per month, adding roughly $25–30/month to a previously-$80–85/month reference setup. Fly.io also spent 2025–2026 deprecating its fixed Launch and Scale plans in favor of pure pay-as-you-go. None of these is a bandwidth-style bill-wide overhaul; each closes one specific gap where something was previously free or under-metered.
Lined up, the four platforms sit on a clear spectrum in the same roughly two-year window: Vercel rewired four separate pieces of its pricing model, Render rewired one, Fly.io added three narrow metered lines without touching its core model, and Railway changed nothing.
What the Cadence Actually Signals
A single repricing is a correction — a company noticed a mispriced dimension and fixed it. Four in eighteen months, each touching a different part of the bill, is a different signal: the underlying cost structure hasn't settled, and pricing is still catching up to it.
Look at what Vercel's four events have in common. Three of the four — granular metering, Fluid Compute, Active CPU pricing — are all about the same underlying problem: making the bill track actual compute consumption more precisely, specifically for spiky, I/O-heavy, often AI-inference-shaped workloads that don't fit neatly into a flat-rate or coarse-metered model. That's not incidental. A platform whose customer base is increasingly running LLM calls and streaming responses through its functions has volatile, hard-to-predict compute costs on its own infrastructure bill — and each Vercel repricing since 2024 has been, in effect, an attempt to pass a more accurate version of that variability on to the customer. The September 2025 seat overhaul is the odd one out structurally, but it's consistent directionally: it's also a move toward charging closer to actual usage (who deploys) rather than a blunt per-head count.
Render's one event and Fly.io's three additions read differently — each closes a specific, identifiable gap (bandwidth was underpriced relative to actual egress cost; snapshot storage was free when it cost Fly.io real money) rather than replacing the pricing model's core unit. Railway's zero events is the cleanest contrast: either its cost base is more predictable, its margins already absorb the variability, or it's simply making a different bet on pricing stability as a retention lever.
None of this makes Vercel's individual pricing changes unreasonable — Active CPU pricing is a genuine cut for the workloads it targets. But "the vendor reprices on a schedule you don't control" is a distinct risk from "the vendor's prices are high." A high price is knowable and budgetable. A pricing model that gets rewritten roughly every four to five months means the spreadsheet you built to forecast next quarter's hosting spend might be modeling the wrong unit of billing by the time the quarter starts.
The Forecasting Cost Nobody Puts in the Comparison Chart
Most hosting comparisons stop at today's sticker price. That misses the actual cost a frequently-repricing vendor imposes: every rewrite means re-deriving your unit economics from scratch. A team that built its Vercel budget model around 2024's per-dimension metering had to redo it for Fluid Compute's time-based billing in February 2025, again for Active CPU pricing's active-millisecond billing in June, and again for September's seat restructuring. Four models in eighteen months isn't four price changes — it's four different spreadsheets.
That risk is structural to any vendor whose margin depends on infrastructure costs it doesn't fully control and workloads it can't fully predict. It's also a risk category a fixed-cost, owned-hardware setup is architecturally immune to — not because the vendor is well-intentioned, but because there's no vendor pricing team deciding when your unit economics change. Compute and bandwidth on a box you own cost what the hardware contract says they cost, on the schedule you signed up for, not the schedule a SaaS pricing team runs its next review on.
Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own, with a Render-compatible API. Compute and bandwidth costs are fixed by your hardware contract, not a quarterly pricing review. Star the repo on GitHub or deploy your first app today.
