Reserved tokens per minute
Your traffic draws on a fixed token budget on the model you chose, so throughput stays flat when another tenant ramps up their batch queue.
Reserved capacity.Promised latency.
Reserve tokens per minute on the model you choose, with a 99.9% uptime SLA. Your traffic never queues behind anyone else's - and anything above your reservation spills onto serverless automatically.
Monthly and annual commitments
Reserved capacity99.9% SLA
DeepSeek V4.1 Flash1M context · reserved
ReservedQwen 3.8 27B256K context · reserved
ReservedOverflowspills to serverless
On demand- Uptime SLA
- 99.9%
- Shortest commitment
- 1 month
- Overflow
- Serverless rates
What you get
Committed throughput, hard SLAs.
Same API as serverless
Burst without dropping
Anything above your reservation spills onto serverless at pay-per-token rates, so a launch never turns into an incident.
Commit, don't lock in
Start monthly, move to annual as volume grows. Discounts scale with the term and the size of the reservation.
Deciding
Two ways to buy the same tokens.
Stay on serverless if
Traffic is still a guess
- Load is spiky, seasonal or not yet predictable
- You are still evaluating which models to standardise on
- You want no forecast and no commitment of any kind
- Monthly volume sits under the break-even point
Move to provisioned if
Latency is in your contract
- P95 latency appears in an SLA you have signed
- Volume is steady and forecastable quarter to quarter
- You need capacity other tenants can't touch
- You want a lower effective rate than on-demand