On a short prompt these two match: $0.10 per million input tokens and $0.50 per million output tokens, with a cache read of $0.01. Haiku 5.5 changes the bill once the input passes 100,000 tokens. The whole request, including cached tokens, moves to $0.50, $2.50, and a $0.05 cache read. Luna’s higher rate starts later, after 272,000 input tokens, at $0.20 and $0.75. A support bot with short prompts will look almost identical. A bot that pastes long transcripts will not.
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.
$0.0003 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
Anthropic's Batch API is half price on input and output. Cache discounts stack with that. Batch jobs are slower, and they suit work that can wait.
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.
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.