GPU sandboxes, by the slice.

Templates include default (the plain CUDA worker) and mujoco; the full list is in the templates docs. A template is an operator-allowlisted image, so a request can only pick a name from that list; GET /gpu/templates lists what this plane serves.

set -euo pipefail
API=https://api.sbx.agent-next.com; H="X-API-Key: $SBX_API_KEY"
ID=$(curl -fsS -X POST -H "$H" -H 'Content-Type: application/json' \
  -d '{"accel": {"vram_gb": 2}}' "$API/sandboxes" |
  python3 -c 'import json,sys; print(json.load(sys.stdin)["sandboxID"])'); test -n "$ID"
curl -fsS -X POST -H "$H" -H 'Content-Type: application/json' \
  -d '{"code": "report[\"gpu\"] = torch.cuda.get_device_name(0)"}' "$API/sandboxes/$ID/exec"
curl -fsS -X DELETE -H "$H" "$API/sandboxes/$ID"   # the reply is the final meter snapshot; billed_s is at its top level

The exec body is Python source run in the lease's worker; torch is pre-bound on the default template and report is returned in the reply. Send cmd instead of code to run a shell command. GPU needs an invited key; the free tier has no GPU.

To plan before you create, read the free seats first, or ask the create to wait for one:

set -euo pipefail
API=https://api.sbx.agent-next.com; H="X-API-Key: $SBX_API_KEY"
curl -fsS -H "$H" "$API/v0/gpu/capacity"   # free seats and the largest slice a create is admitted for
ID=$(curl -fsS -X POST -H "$H" -H 'Content-Type: application/json' \
  -d '{"accel": {"tier": "S", "wait_s": 30}}' "$API/sandboxes" |
  python3 -c 'import json,sys; print(json.load(sys.stdin)["sandboxID"])'); test -n "$ID"
curl -fsS -X DELETE -H "$H" "$API/sandboxes/$ID"   # delete the lease when you are done with it

With wait_s, the create answers 201 as soon as a seat is admitted. If none frees in time it returns the typed gpu_capacity_exhausted refusal with Retry-After, so set your client timeout above wait_s.

Each row says whether the class is live now or on request. A class on request cannot be created yet: asking for it answers the typed gpu_class_unavailable refusal (400) and places nothing. The price in its row is the list price that applies once it is live. Pick a class with accel.class on create. Prices are per second on the exact slice you request, with no minimum billing, and every class sells from 0.5 GiB. Live seat counts need a key: GET /v0/gpu/capacity.

To ask for a class that is not live, say so in the use-case field when you request beta access.

Paused and truly idle time bill 0 seconds on pause-capable templates, unless a lease falls back to freeze-only, which happens when other processes in your lease hold VRAM or the VRAM probe is unreadable: the lease then keeps its seat and its paused time is billed. A template with suspend_mode: "none" (for example mujoco) also bills while it holds the card. The verified tier includes 2 GPU-hours per month and 1 concurrent GPU lease. Card verification is not open yet, so every key is on the free tier for now. The free tier has no GPU: GPU is an invite-only beta, and an invited key gets a one-time 10 GPU-hours.