Priority capacity

Organizations that cannot accept being turned away during peak demand can reserve priority capacity: a committed number of tokens per minute on a model, served ahead of standard traffic. Requests use it by sending service_tier: "priority" or "auto"; see Service tiers.

Sizing a commitment

1. Pull a week of per-minute usage from Usage by model with bucket_width=1h, then a busy day at 1m granularity from the usage report.
2. Take the 95th-percentile minute for input and output tokens separately.
3. Add headroom for growth over the term, typically 20%.

Output tokens are the scarcer resource; commitments price them accordingly.

In the Console

Settings → Billing → Priority capacity shows each commitment, its term, and a utilisation chart with spill-over. Changing a commitment is a contract change; contact your account team.

Watching it

• Priority capacity usage reports per-minute peaks and spilled_requests.
• Spill-over above 1% of priority traffic on busy days means the reserve is too small.
• Utilisation consistently under 50% means you are paying for headroom you do not use.

Billing

Proposed — not offered. Commitment billing terms have not been decided and nothing here is offered.