DeepSeek 网摘Saving on the DeepSeek API: the 50x Cache Gap and Off-Peak Tactics
Two stackable levers hide in DeepSeek's price list: prefix caching and off-peak billing. Used together, the same workload can cost a fraction of the naive price. Prices below are from the official docs verified on October 16, 2026.
Lever one: the 50x cache gap
For Flash off-peak, cached input costs $0.003 per 1M tokens versus $0.15 uncached — a 50x gap. Pro runs $0.022 versus $0.66, a 30x gap. Repeated fixed prefixes — system prompts, knowledge bases, reference docs — cost almost nothing once cached.
Writing for hit rate
- Static first, dynamic last: tool definitions, rules and examples at the top of the prompt; user input and live data at the tail.
- In multi-turn chats, append to earlier messages rather than rewriting them.
- For knowledge-base Q&A, reuse the same document as a stable prefix in batches; Chat Prefix Completion (Beta) helps control the completion boundary.
Lever two: off-peak halves
Peak windows are Mon–Fri 01:00–04:00 and 06:00–10:00 UTC (Chinese public holidays excluded) — that is 09:00–12:00 and 14:00–18:00 Beijing time. Everything else, including full weekends, is off-peak. See the peak/off-peak explainer.
A back-of-envelope example
Say you process 200M input tokens per day (80% cacheable) plus 20M output tokens, all on Flash:
- Peak, no cache: 200M × $0.30 + 20M × $1.20 ≈ $84/day;
- Off-peak, 80% hits: 160M × $0.003 + 40M × $0.15 + 20M × $0.60 ≈ $18.5/day.
About a 4.5x difference, before counting what smarter concurrency scheduling buys you. Flash's 2500 concurrency limit leaves plenty of headroom for overnight batches.
To get started see the Python quickstart; for the full table, the September pricing update.