AI Research Atlas

API context caching on disk

DeepSeek · 2 August 2024

Automatic disk-based prefix caching on the DeepSeek API: cache hits billed at $0.014 per million tokens, up to 90% cheaper than a miss.

Repeated prompt prefixes are cached on disk and served at a fraction of the price with no code change. DeepSeek's cache hit/miss pricing became a template for later cheap long-context agent workloads.

Date
Friday, 2 August 2024
Lab
DeepSeek
Kind
feature
Access
closed API
Price
Cache hit $0.014/M vs miss $0.14/M input (2024-08-02)

Figures

MeasureValueMeasured by
Cache-hit input price$0.014/M tokens
vs $0.14/M for a miss (V2-era pricing)
company

Sources

  1. api-docs.deepseek.com/news/news0802
  2. api-docs.deepseek.com/updates

This record was checked against its sources on 6 October 2026. How we check

Related