How billing works

The basic formula

cost = input tokens × input rate + output tokens × output rate
rates are quoted in USD per million tokens ($/M)

Input and output are priced separately, with output typically 3–5× input. Actual rates are the ones shown on each model's page in the model plaza.

Cache pricing

On models with prompt caching, repeated prefixes — system prompts, long documents, few-shot examples — hit the cache and bill at a lower rate:

So putting stable content at the front of the prompt and variable content at the end cuts cost markedly — most of all in multi-turn chat and document Q&A.

Other billable items

Pay as you go vs. subscription

There are only two things to know:

• Top-ups carry no bonus: pay $100, get $100 in balance. It never expires. Flexible, but no discount.
• Subscriptions give you extra credit: pay $99, get $120 in credit, for example. The credit lands in full at purchase and is valid for 30 days; whatever is left at expiry lapses.

Plan credit vs. balance

An account has two separate pools: the subscription credit (in $, granted in full at purchase and valid for 30 days) and the prepaid balance (your personal balance, or the workspace balance for a team; also in $, never expires). They are never merged or transferred. Web chat and API key calls pick a pool in this order:

• Personal workspace with allowance left on the current plan: billed to the allowance by default. You can switch a message to pay-as-you-go from the billing selector in the composer.
• Allowance used up, plan expired, or no plan: falls back to pay-as-you-go from your balance. A reply that already started stays on the plan; the mode never changes mid-reply.
• Team workspace: there is no personal plan here, so everything is pay-as-you-go from the workspace balance.
• API keys: same order as web chat. In a personal workspace they draw from the subscription allowance first, then fall back to pay-as-you-go from your personal balance; at $0 balance requests return an insufficient-balance error and stop, with no negative balance. Team-workspace keys are always pay-as-you-go from the workspace balance.

What is never billed

• Rejected requests (4xx) — auth failures, bad parameters, rate limits.
• Failed attempts during failover.
• Gateway errors (5xx) that returned no content.
• Models marked free (the “Free” tag in the plaza).

Auditing one call

The response's usage gives the token counts; search the request ID (the X-Maxx-Request-ID response header) in console › Usage › Request log to see exactly what the call cost.

Four ways to spend less

1. Push high-frequency simple tasks down to small models.
2. Move stable prefixes to the front of the prompt to maximise cache hits.
3. Set max_tokens on every call so no single answer runs away.
4. Allow each key only the models it needs, so an anomaly stays inside one workload.