Modal moves coding-agent sandboxes to pay-per-use pricing

A coding agent spends most of its time waiting on the model. It only spikes when it runs an install, a test or a build. On Modal, that pattern is now reflected in the bill. Modal engineer Kevin Yang published the change this week. He calls it active-resource pricing for the Sandboxes.

Modal says a typical Sandbox on its platform uses less than 6 percent of its peak CPU. That is the usage for more than half of its lifetime. The same shape shows up in other common sandbox workloads: reinforcement-learning rollouts, evaluation runs and app preview builders. The agent reasons for stretches of seconds. Then the box spikes for a short compile or test run.

What changed

Every Sandbox on Modal now bursts by default. It starts with a request of 0.125 cores and 128 MiB. The Sandbox can use more CPU and memory than the request when the workload spikes. An optional limit caps how far it can go. It pays each second for what it uses, with the small request as a floor.

A user can still pin a request higher for a guaranteed amount. That reservation is billed even when the Sandbox is idle. A limit caps how far the Sandbox can burst. Modal describes the limit as the size of the computer the Sandbox lives on. For most workloads, the company recommends the default.

The numbers Modal publishes

Modal shows a replay of one real session. A coding agent fixes a vulnerability over 2 minutes 46 seconds. The CPU stays near zero except for brief pytest spikes. Memory sits near 47 MiB. Both stay well under the 0.125-core and 128-MiB request. The agent spends 33 seconds waiting on the model, plus two 60-second waits it asked for.

The company also publishes a worked example. The task is a 20-minute coding agent that installs, reads code, runs tests and finishes with a build. The CPU peaks at 6 cores during install and build. Memory peaks at 2.3 GiB. The reservation needed for that peak would be 1.5 CPU and 3 GiB. That full box would be billed the whole time. Under active-resource pricing, Modal says the same workload bills at 8.4 cents per Sandbox-hour. The reservation would be 20 cents. That is 58 percent less for that one workload, in Modal's own math.

The list rates Modal cites in the post are 14.19 cents per CPU-hour and 2.40 cents per GiB-hour. Modal prices CPU by physical core. It notes that most sandbox providers price by vCPU, and that 1 CPU is about 2 vCPU.

The trade

Bursting is not free of cost. Modal says the saving depends on how spiky the workload is. A long-running web server sits near its peak most of the time. That shape does not benefit the same way a coding agent does. The post includes a calculator. It runs four shapes (coding agent, RL rollout, app preview, web server) against a team's own CPU and memory rates.

Agent sandboxes are now one of the fastest-growing line items in an AI engineering bill. Every parallel agent needs its own isolated environment. A team that sizes those sandboxes for the busiest minute pays for the busiest minute all day. The change at Modal lets the bill follow the shape of the work. The saving on paper is only as good as how closely your workload matches Modal's example. Check your own usage before assuming the 58 percent applies.

Source

Sandbox resourcing: burst, don't reserve, by Kevin Yang at Modal.