Models
Open weights, run on hardware we rent by the hour. Prices are per million tokens.
| id | model | now | context | tok/s | in /M | cached /M | out /M |
|---|---|---|---|---|---|---|---|
ds-flash |
DeepSeek V4 Flash | offline | 256k | 400+ | $0.80 | $0.08 | $1.60 |
kimi-k3 |
Kimi K3 | offline | 1M | 460+ | $5.00 | $0.50 | $25.00 |
Context
The context column is what the running servers actually serve, which can be smaller than what the weights support when a fallback machine is the one live. We quote what you get, not what the model could do on other hardware.
Cached input
Repeated context is billed at the cached rate — that is the same prompt prefix arriving again, which is most of what an agent sends on every turn. The prefix cache is isolated per account: your prefixes are never shared with, or measurable by, another customer.
Pricing example
A representative turn on DeepSeek V4 Flash — 2,000 tokens in, 800 out — costs $0.0029. Only tokens are billed; idle time is not.