Start building
SandBase

Appearance

Pricing and Usage ​

SandBase uses pay-as-you-go billing. Model availability, price formulas, cache multipliers, and capability-specific charges can change independently, so integrations should read live metadata instead of relying on copied price tables.

Manage billing in the Console ​

Use Console → Billing to view your balance, add credits, and manage the billing settings available to your organization. Open Console → Activities → Usage for organization-level request, token, and cost summaries.

Public API boundary

Balance, top-up, billing export, and organization-wide usage endpoints are not currently part of the supported public API. Do not automate Console network requests or depend on internal Console routes.

Read current model pricing ​

First discover an enabled model:

bash
curl 'https://api.sandbase.ai/v1/models?q=gpt' \
  -H "Authorization: Bearer $SANDBASE_API_KEY"

Then request its full model card:

bash
curl https://api.sandbase.ai/v1/models/deepseek/deepseek-v4-flash \
  -H "Authorization: Bearer $SANDBASE_API_KEY"

The list endpoint is optimized for discovery and intentionally omits detailed token, cache, and reasoning prices. Read the selected model's detailed pricing from model_card in the detail response. There is no universal formula that applies to every LLM, media model, embedding model, or API capability.

Check one task's recorded cost ​

When an operation returns a task ID, query its settlement record:

bash
curl https://api.sandbase.ai/v1/tasks/{task_id}/cost \
  -H "Authorization: Bearer $SANDBASE_API_KEY"

The response distinguishes the current cost from estimated_cost and reports whether settlement is final. Do not use cost for billing reconciliation until settled is true. Task records are operational lookup data rather than an accounting ledger. Use the billing records and CSV exports available under Console → Activities → Usage for reconciliation and longer-term records.

Cost-control practices ​

  • Resolve the current model card before launching a new workload.
  • Check capability_tags, supported_modes, and unified_schema; cheaper alternatives are useful only when they support the required input and output behavior.
  • Limit unnecessary prompt context and set an appropriate output-token limit when the selected model supports it.
  • Treat caching as model-specific. Check the model card and returned usage fields instead of assuming a provider-wide discount.
  • Record task IDs only for the task-cost lookup; for asynchronous /v1/run work, record the returned opaque run ID and poll GET /v1/run/{id} only while status is pending or running.
  • Use spending controls exposed in the Console for your organization.

Payment errors ​

An API operation may return HTTP 402 when the organization cannot fund the request. Error envelopes vary by API compatibility surface, so handle the HTTP status and documented error object rather than matching one message string. After resolving billing in the Console, retry according to the operation's idempotency and recovery rules.

Video submissions are also checked against their estimated cost before they start. A single top-up of $10 or more raises your account level, which allows more video tasks to run at the same time. See Concurrency limits.

See Models and Pricing, Task Cost, and Errors for the current public contracts.