# Modal vs Prime Sandboxes: sandboxes for agent training Modal and Prime both court reinforcement learning teams. The 2 vCPU, 4 GiB benchmark costs $3.97 on Modal and $1.52 on Prime Sandboxes. **Runtime costs 84% less than Modal and 58% less than Prime for rollouts that wait on a model.** In the benchmark, 1,000 one-minute runs of a 2 vCPU, 4 GiB sandbox with 20 CPU-seconds of work each, Modal comes to $3.97, Prime to $1.52 and Runtime to $0.64. ## At a glance | | Modal Sandboxes | Prime Sandboxes | Runtime | | --------- | -------------------------------------------------------- | ---------------------------------------------- | -------------------------------------------------------- | | Isolation | gVisor, a user-space kernel | Hardware-virtualized microVM, own guest kernel | Firecracker microVM, own kernel | | CPU | $0.1419 per physical core-hour (2 vCPUs), request or use | $0.02 per vCPU-hour while running | $0.025 per vCPU-hour of measured CPU, with a small floor | | Memory | $0.0240 per GiB-hour, request or use | $0.0125 per GiB-hour | $0.0075 per reserved GiB-hour | | Snapshots | Carry state past a session | Listed as coming soon | Pause and forks keep memory and processes | | GPUs | Yes | Listed as coming soon | No | | At once | Set by plan | 1,024 per account to start | 100 paid, after 50 in the first week | ## Training stacks Modal's argument is one platform: GPU trainers and inference as functions, environments as sandboxes beside them. Prime's is the environment itself at scale and at a low price, microVMs by the thousand on its own compute. Runtime is the environment layer too, billed by use, which suits rollouts that spend most of each step waiting for a policy model to answer. ## Cost for the same job The benchmark in full, 60 seconds a run (one physical core on Modal, Prime's 5 GiB disk): ``` Modal CPU 1,000 × 60 s × 1 core × $0.00003942 = $2.37 Memory 1,000 × 60 s × 4 × $0.00000667 = $1.60 Total $3.97 Prime CPU 1,000 × 60 s / 3,600 × 2 × $0.02 = $0.67 Memory 1,000 × 60 s / 3,600 × 4 × $0.0125 = $0.83 Disk 1,000 × 60 s / 3,600 × 5 × $0.0002 = $0.02 Total $1.52 Runtime CPU 1,000 × 20 s / 3,600 × $0.025 = $0.14 Memory 1,000 × 60 s / 3,600 × 4 × $0.0075 = $0.50 Total $0.64 ``` For a month of 100,000 runs: $396.60 on Modal, $151.67 on Prime, $63.89 on Runtime. When both CPUs stay busy for the whole minute, it is $3.97 on Modal, $1.52 on Prime and $1.33 on Runtime. ## Which one to pick - **Modal** when environments and GPU training should live in one platform. - **Prime Sandboxes** for a thousand microVMs at once inside Prime Intellect. - **Runtime** for the lowest price per rollout, forks of a warmed-up environment today, and [Harbor](/docs/harbor) and [Inspect](/docs/inspect) integrations for eval suites. See [Runtime vs Modal](/docs/modal-sandbox-alternative) and [Runtime vs Prime Sandboxes](/docs/prime-sandboxes-alternative). ## Sources Modal rates checked 23 September 2026; Prime rates checked 25 September 2026. - [Modal pricing](https://modal.com/pricing), [Modal resources and billing](https://modal.com/docs/guide/resources) - [Prime Sandboxes overview](https://docs.primeintellect.ai/sandboxes/overview), [Prime Sandboxes launch](https://www.primeintellect.ai/blog/sandboxes) - Runtime [pricing](/docs/pricing) Facts on this page were checked on 25 September 2026.