What prompt caching actually costs, and where the tenth does not hold
Caching pays off when the same prefix goes out over and over: a long system prompt, a document, a conversation history.
At most providers a cache read costs about a tenth of the input price[1,2,3,7,8,9,10]. OpenAI is a tenth on every GPT-6 and GPT-5.6 model, and Gemini 3.8 Flash is $0.075 against $0.75[2,3].
Mistral takes 90 percent off cached input, Alibaba bills a cache hit at 10 percent of input, and Moonshot is $0.30 against $3 on kimi-k3[7,8,9].
Sakana's fugu-ultra-v2.0 is $0.50 against $5, while fugu-max-v1.0 is $0.25 against $2, an eighth[10].
Some sit far below that[1,4,6]. Claude Fable 5.1 and Claude Mythos 5.1 read at 0.025 times the input price, a fortieth, where every other Claude model is a tenth[1].
DeepSeek, off-peak, is $0.003 against $0.15 on V4.1 Flash, a fiftieth, and $0.022 against $0.66 on V4 Pro, a thirtieth[4]. Meta's contributor tier is $0.002 against $0.10, also a fiftieth[6].
xAI goes the other way. grok-4.6 charges $0.50 against a $2 input, a quarter[5].
Some providers also charge for writing to the cache[1,2,8]. Claude charges 1.25 times the input price for a 5 minute cache and 2 times for a 1 hour cache[1]. OpenAI lists cache writes at 1.25 times input, $5 against $4 on gpt-5.6-sol[2].
Alibaba bills creating an explicit cache at 125 percent of input[8]. Gemini charges for keeping a cache by the hour, $0.50 per million tokens per hour on 3.8 Flash[3].
Where writes are charged, a prefix that is written and never read again costs more than not caching at all. On Claude a 5 minute cache pays for itself after one read, and a 1 hour cache after two[1].