DeepSeek cached input pricing explained
A low input price is not the same as a low bill. If your prompts reuse long prefixes, most of your prompt tokens can be cache hits, and the cached rate becomes the number that decides your cost.
The three lines that matter
- Fresh input: prompt tokens that were not in the cache.
- Cached input: prompt tokens that hit the cache.
- Output: generated tokens.
Compare a provider on all three. A provider with a lower fresh input price can still be more expensive if your workload is mostly cached.
Worked example
Take 1,000,000 prompt tokens per month, 80 percent of them cached, plus 100,000 output tokens.
At Fikra DeepSeek V4.1 Flash rates:
- Fresh input: 200,000 / 1M * $0.05 = $0.01
- Cached input: 800,000 / 1M * $0.0075 = $0.006
- Output: 100,000 / 1M * $0.15 = $0.015
- Total: $0.031
At SiliconFlow DeepSeek-V4.1-Flash rates:
- Fresh input: 200,000 / 1M * $0.15 = $0.03
- Cached input: 800,000 / 1M * $0.003 = $0.0024
- Output: 100,000 / 1M * $0.60 = $0.06
- Total: $0.0924
In this example Fikra is lower overall because output and fresh input dominate. Change the mix to 95 percent cached and a small output share, and the cached rate can flip the result. Run your own numbers.
How to measure your cache hit rate
- Read
prompt_tokens_details.cached_tokensfrom the response usage block. - Divide cached tokens by prompt tokens.
- Split your traffic by workload: coding agents, chat, batch, retrieval.
What Fikra does
Fikra bills cached tokens at the cached rate and shows cached tokens in the usage response. The rate is set from one config source and published on /pricing. Cached input can never cost more than fresh input.
Related: /compare/siliconflow-alternative, /pricing, /stats, /free-deepseek-api.