GPT-5 mini vs GPT-5.6 Luna: cost compared
Last verified
GPT-5.6 Luna is 20% cheaper on input and 40% cheaper on output than GPT-5 mini. Whether that makes it the right choice depends entirely on whether it does your task well enough — this page gives you the cost side of that trade, precisely.
Rates side by side
| $ per million tokens | GPT-5 mini | GPT-5.6 Luna |
|---|---|---|
| Input | $0.25 | $0.20 |
| Output | $2.00 | $1.20 |
| Cached input | $0.03 | $0.02 |
| Output : input ratio | 8.0× | 6.0× |
What the difference is worth
Monthly cost for the same workload on each, without caching:
| Workload | GPT-5 mini | GPT-5.6 Luna | Difference |
|---|---|---|---|
| Support chatbot 1,000 conversations/month · 2,000 in + 500 out each | $1.50 | $1.00 | $0.5000 saved |
| RAG search 10,000 queries/month · 8,000 in + 400 out each | $28.00 | $20.80 | $7.20 saved |
| Bulk extraction 10,000 documents · 3,000 in + 300 out each | $13.50 | $9.60 | $3.90 saved |
| Coding agent 100 runs/month · 400,000 in + 20,000 out each | $14.00 | $10.40 | $3.60 saved |
Which to use
GPT-5.6 Luna if the task is well-specified and mechanical — extraction, classification, formatting, routine transformation. The saving is real and compounds with volume.
GPT-5 mini if the task involves genuine reasoning, long-horizon planning, or work where a wrong answer is expensive to catch. Paying 1.3× more on input is trivial compared to the cost of shipping a bad result.
Both are OpenAI models, so switching is usually a one-line change and the tokenizer is the same — which makes this a genuinely easy experiment to run. Test on your own workload before deciding.
Don't skip caching
Cached input costs $0.03 on GPT-5 mini and $0.02 on GPT-5.6 Luna. If your prompts share a stable prefix, caching on the more expensive model can beat switching to the cheaper one outright — worth checking before you migrate anything.
Sources
OpenAI rates verified 2026-08-05 (source). OpenAI rates verified 2026-08-05 (source).