Launch: 8 Burstable vCPUs Everywhere, and You Only Pay for Active Seconds

Starting today, every sandbox and app on PandaStack runs on a new compute layer, and the bill meters what you use instead of what you reserve. Here's what changed, in the order you'll notice it.

8 burstable vCPUs, everywhere

Every first-party template now boots with 8 vCPUs — sandboxes, code interpreters, browsers, apps. On a quiet host, a single sandbox can saturate all 8 cores, so installs, builds, and test runs get real parallelism instead of a small fixed slice. Under contention, the kernel fair-shares cores evenly across busy guests: burst is free headroom when the host is quiet and an equal share when it isn't — never a tax on your neighbors, and never theirs on you.

Pay for active seconds, not committed hours

CPU is now billed on active CPU-seconds — the compute you actually burn. In production measurement, a VM running a sustained build was billed 160 CPU-seconds over a window where committed-basis billing would have charged 1240, and an idle-but-running VM accrued just 2 CPU-seconds across 306 seconds of wall clock. Memory is billed on your working set — resident memory, clamped to your committed size — so the meter can only read at or below the committed basis it replaces. Idle sandboxes now cost roughly what they consume: almost nothing.

Apps: opt-in autoscaling

Apps can now opt in to horizontal scaling: set max_instances (2 or 3), and when average CPU or memory crosses 80% we start a warm instance from your baked deploy image — measured at 21-25 seconds to serving — and load-balance traffic across instances. When traffic falls off, extra instances scale back in on their own. Nothing changes unless you opt in.

Idle means nearly free

Idle apps already sleep and wake on traffic — the first request after a wake serves a real response, not an error page. Managed databases auto-suspend after 15 minutes idle and wake on connect. With active-CPU and working-set metering on top, a project that's quiet overnight now bills like a project that's quiet overnight.

What didn't change

Databases. Managed Postgres remains the guaranteed class: it pays its reservation, is never overcommitted, and none of the new burst or squeeze machinery touches it. Your database's RAM is its cache, and we leave it alone.

Full rates are on the pricing page, and the engineering deep-dive — the fair-share measurements, the pressure ladder, and the honest limits — is in "Tidal: turning the host into a cache." Everything above is live now for every account: no action needed, no migration, and the new meters measure at or below the committed basis they replace.