Pay as you go
Clear compute rates for training and usage-based serverless inference. Add credits when you need them and review your estimate before you start.
Training compute
| Hardware | Memory | Per minute | Per hour |
|---|---|---|---|
| NVIDIA A100 | 80 GB | $0.06650 | $3.99 |
| NVIDIA H100 | 80 GB | $0.10817 | $6.49 |
| NVIDIA B200 | 180 GB | $0.18317 | $10.99 |
Serverless inference
| Serving profile | Hardware | Per second | Per hour |
|---|---|---|---|
| Qwen3 8B | A40 / RTX A6000 · 48 GB | $0.0018 | $6.48 |
Sandboxes
Run agents, code, and tools in isolated cloud workspaces. Contact us about the compute and storage configuration for your workload.
Discuss sandbox capacity →Hosted model APIs
Calls to external models use the selected model’s published input, cached input, and output token rates. Review model calls, search, and retained storage separately from compute.
Read the documentation →