Off peak, DeepSeek Flash input is $0.15 per million tokens on a cache miss, against Luna’s $0.10. Flash output is $0.60, against Luna’s $0.50. Flash’s cache hit is $0.003, cheaper than Luna’s $0.01 cache read. The larger difference is the clock. DeepSeek’s peak window is double, on weekdays from 01:00 to 04:00 and 06:00 to 10:00 UTC, except Chinese public holidays. Luna does not publish a peak rate. DeepSeek’s English API docs did not list a batch discount, so this comparison does not invent one.
For 1,000 input tokens, 400 output tokens, and 50 requests a day, in US dollars: GPT-6 Luna is the cheapest of these, at $0.45 a month.
$0.0003 for one request, and $0.02 a day at 50 requests a day.
Show the maths
New input: 1,000 tokens × $0.10 per million = $0.0001
Output: 400 tokens × $0.50 per million = $0.0002
1,500 requests a month × $0.0003 = $0.45
Caching half the input would save about $0.07 a month
That uses the published cache-read price of $0.01 per million tokens. It assumes half of these input tokens are already in a cache. Writing the cache can cost extra, and that write price is not included.
Batch processing would save about $0.23 a month
OpenAI's batch table lists half the standard input and output price for this model. Batch jobs are slower, and they suit work that can wait.
GPT-4.1 nano would cost about $0.06 less a month
It is a cheaper model in the same broad class on this price list. It is still a different model, and the replies will not match.
A reply half as long would save about $0.15 a month
Output tokens are often the expensive part. Asking for a shorter answer cuts that cost directly.
$0.0004 for one request, and $0.02 a day at 50 requests a day.
Show the maths
New input: 1,000 tokens × $0.15 per million = $0.0002
Output: 400 tokens × $0.60 per million = $0.0002
1,500 requests a month × $0.0004 = $0.59
Caching half the input would save about $0.11 a month
That uses the published cache-read price of $0.003 per million tokens. It assumes half of these input tokens are already in a cache. Writing the cache can cost extra, and that write price is not included.
A reply half as long would save about $0.18 a month
Output tokens are often the expensive part. Asking for a shorter answer cuts that cost directly.
Cached prices here are read prices. Writing a cache can cost extra, and that write price is on the provider’s page, not in this total. A US-only endpoint is 1.1× at some providers and is not included. Fast modes are not included.