I host a light-weight project for a few thousand users on Vercel. It's a small Express.js app with a few hundred uncached requests per hour, mostly I/O-bound (API calls, Postgres).
Two weeks ago I moved it from Node.js runtime to the new Container runtime. Since this new runtime also uses Fluid compute, I didn't expect significant changes in cost:
Containers deploy as Vercel Functions on Fluid compute, scaling with traffic and using Active CPU pricing, which doesn't charge for idle time spent waiting for API responses, database queries, or LLM inferencing.
The changelog says the same: "runs on Vercel Functions, and is billed with Active CPU pricing". Our spending went up a lot anyway.
Before: Node.js function (Oct 28–29)

CPU and memory track invocations closely. Baseline is ~10–20 s of Active CPU and ~0.2 h of Provisioned Memory per hour. The spikes are a cache-refresh job.
After: Container runtime (Nov 15–16)

Invocations are roughly the same, but:
- Provisioned Memory is at ~6–8 h per hour, about 30x more, and it no longer follows traffic at all. The refresh spikes don't even show up.
- Active CPU has a floor of ~2 min per hour, about 10x more.
Why?
The reason seems to be this behaviour mentioned in the Container runtime docs:
Functions not receiving any traffic for 5 minutes in production environments, and 30 seconds in preview environments, will automatically scale down.
While regular Fluid functions are paused as soon as the last in-flight request finishes, a container instance stays alive for 5 minutes. At one request every few seconds, containers are never stopped. So containers are effectively billed 24/7.
Additionally, the memory numbers above mean multiple instances (with 2 GB each) are constantly alive. So apparently optimized concurrency (several requests sharing one instance) is not supported for containers.
So for low-but-steady traffic, the Container runtime is much more expensive. This surprised me, so I thought I'd warn others about it.
Questions
- Is this the expected billing behavior for containers, or am I missing something?
- Is there anything I can configure to reduce cost?
- How did migrating to Containers runtime affect billing in other projects?