Cached input is now priced much closer to the real upstream rate. For every DeepSeek route, the cached-input price is set at the 20% margin floor, and the public pricing table reads from that one config source.
What changed
| Model | Upstream cached | Retail cached | Margin |
|---|---|---|---|
| deepseek-v4-flash-0731 | $0.0090 | $0.01125 | 20% |
| deepseek-v4-flash-ark | $0.0140 | $0.01750 | 20% |
| deepseek-v4-pro-ark | $0.0440 | $0.05500 | 20% |
| deepseek-v4.1-flash | $0.0060 | $0.00750 | 20% |
| deepseek-v4.1-flash-1017 | $0.0060 | $0.00750 | 20% |
| deepseek-v4.1-flash-ark | $0.0030 | $0.00375 | 20% |
Repeated prompt prefixes are billed at the cached rate. Cached input can never cost more than fresh input. If you run coding agents or multi-turn tools, this is the line that matters most.
Public stats
/stats is now live. It publishes aggregate volume, reliability from our own probe and live traffic, latency for cold and cached requests, cache hit rate per model, and a model integrity table. Empty states are explicit. No customer identifiers and no per-user numbers appear there.
Free DeepSeek API
The free DeepSeek API page now shows live availability from the same server-side gate that routes the model, the exact model ID, the free limits, expiry conditions and copy-paste setup for Cline, Roo Code, Open WebUI and the OpenAI SDKs.
Setup guides
The new guides hub has verified setup pages for coding agents and OpenAI-compatible clients.