Pay for tokens. Nothing else.

No subscription, no seat licences, no minimum spend, no annual commitment. You fund a balance and each request draws from it at the model's published rate.

ModelInputOutputContext
gpt-4.1 $2.00 / 1M tokens $8.00 / 1M tokens 1M

Rates are the provider's list price, passed through. More models are being added — the catalog in the console is always the authoritative list, and it is what your requests are priced against.

How a charge is worked out

Nothing here is estimated after the fact. Each step below happens on the request itself.

1  Reserve

Before any provider is contacted, the worst-case cost of the request is held against your balance. If you cannot cover it, the request is refused then — not after the tokens are spent.

2  Serve

The request runs. Tokens in and out are counted from the provider's own usage report, not estimated from the text.

3  Settle

The reservation is replaced by the real amount and the difference returns to your balance. A request that never reached a provider costs nothing at all.

Questions people actually ask

What if a stream dies halfway?

You are charged for what the provider reported and nothing more. A stream that fails after the first byte is recorded as a failed request, and its held amount is reconciled rather than kept.

Can I go negative?

No. The reservation happens before the provider call, so the balance is the ceiling. There is no invoice arriving later.

Do unused credits expire?

No. A funded balance stays funded until you spend it.

Is there a free tier?

Not today. You pay for what you use from the first request, at the rates above.

How precise is the accounting?

Amounts are held in micro-dollars — millionths of a dollar — so a request costing a fraction of a cent is recorded exactly rather than rounded to zero.

Can I cap a single key?

Keys carry scopes today. Per-key spend limits are on the roadmap; until then the balance is the shared ceiling for the workspace.