Inference Billing and Availability

You pay for GPU runtime

An inference instance is billed from your prepaid credit for the GPU time it uses, plus the platform margin. There is no per-token and no per-page charge, and no weekly budget or schedule as on coding instances.

  • Credit is deducted every 15 minutes while the instance runs, at any time of day.
  • You need at least $5.00 of credit to create an instance.
  • Price ceiling. You set the most you are willing to pay per hour. Hardware above it is not offered, and an instance that can only be provisioned above it does not start.
  • At zero credit the instance stops. Requests then get HTTP 402 with the code credit_exhausted until you top up and restart. Turn on auto top-up to avoid it.

Why this is not always cheaper than per-token pricing

A GPU costs the same per hour whether it handles one request or thousands. That makes it cheap per request when the load is steady or high, and expensive per request when the load is small. A pay-per-token API charges for exactly what you use.

Roughly: if your GPU would be busy for most of the hours it runs, a GPU instance is usually cheaper. If it would idle most of the time, an API is cheaper unless you pause the instance between uses. The cost calculator shows the break-even for your workload.

Idle pause and wake on request

You can let the instance pause itself when nobody uses it. Choose the idle time in the wizard: 15 minutes, 30 minutes, 1 hour, 2 hours or 12 hours without requests. Auto-stop is off by default, and the wizard asks you to choose.

  • A paused instance stops the GPU charge. The disk with the model stays, and its small storage charge continues while the instance is paused. A restart after a pause does not download the model again.
  • An authenticated request wakes it. The first request gets a 503 with Retry-After while it starts. See Inference Limits.
  • The wizard and the instance page show how long a restart takes: a restart after a pause is shorter than a restart on a new host, and both grow with the model size.

Availability is best effort

Instances run on rented GPUs from GPU marketplaces and cloud providers. A host can be interrupted or disappear. When that happens syndicAI provisions a new host and the instance comes back, and its restart time is the "new host" estimate shown in the wizard. There is no SLA. Do not use an inference instance as the only path for something that must never fail: retry on 503, keep a fallback, and test your client against a restart.

Hardware from verified hosts is the default. Community hardware is available as an option; do not choose it for sensitive documents.

Deleting an instance

Deleting stops the endpoint and revokes its keys at once, and releases its name. Storage and GPU charges stop with it.