dssubWhat is a cache hit?
A provider can reuse earlier computation when your request matches a stored input prefix. Shared instructions or conversation history are common examples. The model still generates a new answer.
A cache hit requires a stored prefix that satisfies the provider’s matching rules.
How the price works
Matched input uses the cache-read rate. Unmatched input uses the regular input rate; generated tokens use the output rate.
10,000 input tokens: 8,000 cached + 2,000 uncached.
At $0.02 / 1M cached and $0.20 / 1M uncached, input costs $0.00056: (8,000 × $0.02 + 2,000 × $0.20) ÷ 1,000,000. Output is billed separately.
Why a request can miss
A changed prefix, an unavailable cache entry or a different provider route can prevent reuse. Cache behavior varies across services. DeepSeek describes its cache as best effort; inspect reported cache-hit and cache-miss usage instead of assuming every repeat will qualify.
For our subscription comparison, cached input, uncached input and output all consume the shared allowance at 1×. The displayed subscription price assumes full allowance use.
Cache and FP8 can work together
Caching reuses input computation. FP8 describes numerical representation during inference. An FP8 route can have cache hits and misses. Read About FP8 →
Further reading
Independent explainer · Reviewed November 5, 2026.