Contents
At a glance
- Model ID
- gpt-5.6-luna
- Input price
- $0.20 per 1M tokens (checked August 2026)
- Cached input price
- $0.02 per 1M tokens
- Output price
- $1.20 per 1M tokens
- Context window
- 1,050,000 tokens
- Max output
- 128,000 tokens
- Positioned as
- Optimised for cost-sensitive workloads
- Versus Sol
- One twenty-fifth of the price on both input and output
What GPT-5.6 Luna is
gpt-5.6-luna is OpenAI's cost-optimised model ID in the GPT-5.6 family, described as optimised for cost-sensitive workloads. It carries the same 1,050,000-token context window and 128,000 maximum output tokens as gpt-5.6-sol and gpt-5.6-terra.
That combination - frontier-sized window, cheap tokens - is what makes this model ID interesting for automation rather than for chat. High-volume, high-token, low-difficulty work is exactly the shape that used to be uneconomic.
What it costs
OpenAI's pricing page in August 2026 lists gpt-5.6-luna at $0.20 per million input tokens, $0.02 per million cached input tokens and $1.20 per million output tokens.
That is one twenty-fifth of Sol on both meters. Put concretely: a million tokens of input costs 20 cents. Classifying 100,000 support emails at 500 tokens each is 50 million input tokens, which is $10.00, plus whatever the labels cost to emit - a few dollars more. The same job on the frontier tier is $250.
The jobs it does as well as the frontier
Be specific about this rather than trusting a benchmark. The tasks where a small model routinely matches a large one are the ones with a narrow output space and clear instructions: classification into a fixed set of labels, extraction into a defined schema, routing, deduplication, sentiment, language detection, converting messy input into structured JSON, and short summarisation of a single document.
The tasks where it does not are the ones with an open output space and multiple valid paths: multi-step planning, tool selection among many similar tools, code changes across files, and any reasoning chain where an early error compounds. See planner-executor for the pattern that lets you use both - an expensive model to plan, a cheap one to execute each step.
What you should know
The cost difference is so large that it changes what is worth building. Work that is uneconomic at $5 per million input becomes routine at 20 cents: enriching every record rather than a sample, classifying an entire archive, running a second model as a checker on the first model's output. That last one is a genuinely good use of a cheap tier - a guardrail pass that costs a fortieth of the main call is close to free insurance.
Verify current prices before budgeting. These figures were checked in August 2026 against OpenAI's pricing page, and this tier of the market reprices frequently.
GPT-5.6 Luna FAQ
How much does gpt-5.6-luna cost?
OpenAI listed $0.20 per million input tokens, $0.02 per million cached input tokens and $1.20 per million output tokens in August 2026. That is one twenty-fifth of the frontier tier on both meters.
What is GPT-5.6 Luna good at?
Work with a narrow output space and clear instructions: classification into fixed labels, extraction into a defined schema, routing, deduplication, language detection, converting messy input into structured JSON, and short single-document summarisation. It matches larger models on these routinely.
Where does Luna fall short?
Open-ended work where several paths are valid and an early mistake compounds - multi-step planning, choosing between many similar tools, multi-file code changes, and long reasoning chains. A planner-executor split lets you use an expensive model to plan and Luna to execute each step.
Does Luna have a smaller context window?
No. It carries the same 1,050,000-token context window and 128,000-token output ceiling as Sol and Terra. Cheap tokens plus a frontier-sized window is what makes it useful for bulk document work rather than only for chat.
Can I use a cheap model to check an expensive one?
Yes, and it is one of the better uses of this tier. A guardrail or verification pass costing a fortieth of the main call is close to free insurance against a bad output reaching a customer. Score it on your own eval set to confirm the checker actually catches what you care about.
What does it cost to classify 100,000 emails?
At 500 input tokens each that is 50 million input tokens, which is $10.00 at the $0.20 per million rate, plus a few dollars for the short labels emitted. The same job on the frontier tier would be around $250 of input alone.
Primary source
Checked August 2026. Anything priced or rate-limited moves, so verify against the source before you budget.
Want this built rather than explained?
Knowing what gpt-5.6 luna means is one thing; having it running in your business is another. We scope one workflow at a time, quote a fixed price in writing after a free consult, build it, and hand it over with documentation you own.
