Pricing and limits

Ampersand uses token-based pricing. All prices are in US dollars per one million tokens, for the public model ampersand-v0.1.

Cached input is the input that the model server reports as reused from its cache. It is charged once, at the cached-input price. Cache availability depends on the model server and the request; a cache hit is not guaranteed.

Usage includes the model calls needed to complete your request, including continuations after tool results and context compaction. Reasoning tokens are included in output usage, even though private reasoning text is not returned. Your application runs your custom tools and pays any costs for those tools.

For example, 1 million uncached input tokens, 1 million cached input tokens, and 1 million output tokens cost $6.40.

Each execution is priced when it starts. Executions already running keep their prices. Queued work and later continuations use the prices in effect when they start. View your balance and settled usage in the developer console.

Account tiers

The Normal and High tiers use the same token prices. A tier changes request limits, concurrent execution limits, and session limits. All API keys for an account share these limits.

Work requests create sessions or submit events other than cancellation-only requests. They count against both the API request limit and the work request limit. A burst allows short periods of faster requests within the account's rate allowance. The session count does not reset over time. Deleting a session does not free a slot. An explicit account capacity setting can override the concurrent execution limit. Other limits, including available credit and service capacity, can also prevent a request from starting.

See the API reference for request errors and retry behavior.