Artificial intelligence news.
Straight from the original announcements.

Modal introduces burst-based pricing for Sandbox workloads

November 12, 2026By the AI News Bites Editorial TeamSource: Modal
Modal introduces burst-based pricing for Sandbox workloads
Image: Modal

Modal is changing how its Sandboxes use and are billed for computing resources: they burst beyond a small default request when workloads spike, then release those resources when demand falls. The company says this active-resource pricing is a better fit for coding agents, RL rollouts, evals and app builders.

Key points
  • Sandboxes burst CPU and memory beyond a small default request.
  • Billing is based on active resource use each second.
  • The default request is 0.125 cores and 128 MiB.
  • Modal’s example reports 58% savings versus a fixed reservation.
  • Memory limits can cause eviction when a Sandbox exceeds its request.

Pricing follows usage instead of peak capacity

Modal says a typical Sandbox spends more than half its life using less than 6% of its peak CPU. In a coding-agent workload, for example, the Sandbox may remain nearly idle while the model reasons, then require more capacity to install dependencies, run tests or build a project.

Under the company’s active-resource pricing model, every Sandbox starts with a default request of 0.125 cores and 128 MiB. It can grow as needed, and customers pay each second for the higher of the request and the resources actually used. Customers can request more capacity for a guaranteed baseline, but they pay for those resources while the Sandbox is idle as well.

Modal recommends estimating costs from average usage rather than peak usage. Its formulas are CPU cost equal to runtime multiplied by the CPU rate and average cores used, and memory cost equal to runtime multiplied by the memory rate and average GiB used.

A sample coding task shows lower reported costs

Modal’s illustrative 20-minute coding-agent example includes installation, reading, editing, testing, type-checking and a final build. The workload averages 0.44 CPU and 0.94 GiB of memory, while its peak use reaches 6.00 CPU and 2.30 GiB.

Using Modal’s listed rates of 14.19¢ per CPU-hour and 2.40¢ per GiB-hour, the example costs 8.4¢ per Sandbox-hour under bursting. A fixed reservation sized at 1.5 CPU and 3 GiB costs 20.0¢ per Sandbox-hour in the comparison. Modal describes this as 58% savings versus the fixed-size reservation.

The company says the comparison is illustrative rather than based on production traces. It also notes that 1 CPU is approximately 2 vCPU, and that Modal prices CPU by physical core while most Sandbox providers price by vCPU.

Defaults, limits and the risk of memory eviction

Modal says most workloads do not need a custom request or limit. By default, CPU can burst up to 16 cores above the request, while memory has no limit. A limit can still help keep benchmark runs consistent or cap costs from an uncapped process or runaway agent. For example, the company shows a Sandbox configured with a 0.125-core request, a 4.0-core limit, a 128 MiB memory request and a 2048 MiB memory limit.

CPU is guaranteed at the requested level, although a bursting Sandbox may slow down if a machine runs out of available CPU. Memory has a more serious failure mode: a Sandbox using more than its request may be evicted with an out-of-memory error if the machine runs out of memory. Modal says Sandboxes under 8 GiB have an OOM rate below 0.0001% per minute, or about 1 eviction per 694 Sandbox-days, and says it has no SLO yet for larger Sandboxes.

To reduce the impact, Modal points to exit snapshots, filesystem snapshots, Volumes and termination-reason tools for detecting evictions and retrying. Customers whose workloads cannot tolerate OOMs at the stated rates can set a higher memory request, then use the dashboard’s Exit Reasons chart to adjust it.

Related news

AWS shows how to route coding agent requests with NeMo Switchyard
For developersNovember 13, 2026, 23:30

AWS shows how to route coding agent requests with NeMo Switchyard

AWS has published a sample deployment that routes coding agent requests between models on Amazon Bedrock using NVIDIA NeMo Switchyard and LiteLLM. The approach measures…

AWS Says Agents Wrote Its 200+ Service Emulator Without Human Code Review
For developersNovember 12, 2026, 22:52

AWS Says Agents Wrote Its 200+ Service Emulator Without Human Code Review

AWS says autonomous agents wrote and tested every line of AWS Emulator, a free local copy of more than 200 AWS services, while a merge gate controlled which changes…

Amazon says autonomous AI agents built its AWS Emulator
For developersNovember 11, 2026, 18:25

Amazon says autonomous AI agents built its AWS Emulator

Amazon has released AWS Emulator, a local emulator of AWS services that the company says was written entirely by autonomous AI agents. Thousands of agents write tests…