What You Can Do with GPT-6 Luna
GPT-6 Luna is the fast, low-cost tier of the GPT-6 family. Four things about it change how you design around it, and none of them are in the marketing copy.
Output Costs 5x What Input Costs
$0.03 per 1M input against $0.15 per 1M output. Long prompts are close to free; long answers are not. Capping max completion tokens moves your bill far more than trimming the prompt does.
A 1M Window With a Cliff at 272K
The context window is 1,050,000 tokens, but crossing 272K input tokens reprices the entire request at 2x input and 1.5x output - not just the excess. A 280K-token prompt costs more than two 140K ones.
It Reasons Before It Answers
In our runs the first streamed token took 5.0 to 5.6 seconds of a 7.2 to 7.4 second response. Streaming does not hide that gap. If you need text on screen sooner, lower the reasoning effort rather than the output cap.
Structured Output Holds Without a Schema
Asked in plain prose for four named fields as JSON, it returned exactly those four keys in a fenced block, with no schema, no function definition, and no retry. The sample above is that unedited response.
GPT-6 Luna Pricing
Input Tokens
$0.03 per 1M tokens, against OpenAI's official $0.10.
Output Tokens
$0.15 per 1M tokens, against OpenAI's official $0.50.
Cached Input Reads
$0.003 per 1M tokens, against OpenAI's official $0.01.
Long-Context Tier
Requests above 272K input tokens follow OpenAI's own surcharge: 2x on input and 1.5x on output.
GPT-6 Luna Use Cases
Support Triage at $0.05 per 1,000 Tickets
Summary plus severity, product area, affected version, and a priority label, in one call per ticket. A 50,000-ticket backlog reprocesses for about $2.30.
Batch Enrichment With a Cached Prefix
Batch jobs repeat the same instructions on every row. Keep that prefix byte-identical and the repeated part reads at $0.003 per 1M instead of $0.03 - a tenth of the input cost across the whole run.
Default Tier Behind a Router
Send everything to Luna and escalate only what fails a check. Because Sol costs roughly 20x, a router that keeps 95% of traffic on Luna pays about a tenth of an all-Sol pipeline - the 5% that escalates is half the remaining bill.
Whole Documents Under the 272K Line
Reports, transcripts, and contracts fit the 1M window, but stay under 272K input tokens per request to avoid the 2x repricing. Chunking below that line is usually cheaper than one large call.
How to Use GPT-6 Luna
Luna is a volume model, so test it the way you will run it: on a real record, with the settings you can afford at scale.
Paste One Real Record
Use an actual ticket, transcript, or document rather than a toy prompt. Luna's failure mode is messy input, not hard reasoning.
Tune Down, Not Up
Start at low reasoning effort and a tight max completion tokens. Output is 5x the price of input, so the output cap is the lever that decides your bill.
Multiply the Token Count
Take the usage reported under the response, multiply by your daily volume, and you have the monthly cost before you write any integration code.
Build with the GPT-6 Luna API
What matters in a batch integration is not the first call but the ten-thousandth: cost per record, cache behaviour, and what happens when one row fails.
Drop-In Model ID
Point an existing OpenAI-compatible client at gpt-6-luna and change nothing else. Streaming, system prompts, and JSON-shaped output all behave as they do upstream.
Cached Prefixes Pay Off Here
Batch jobs repeat the same instructions on every row. Cached reads cost $0.003 per 1M against $0.03 for fresh input, so a stable prompt prefix is worth structuring for.
Per-Response Cost Accounting
Every response reports input, output, cached reads, and credits consumed, so a pipeline can log unit economics per record instead of discovering them on the invoice.
Escalate to GPT-6 Sol Without Re-Plumbing
Sol and Luna share one endpoint, one key, and one balance, so routing a hard row to the flagship tier is a model-ID swap, not a second integration.
GPT-6 Luna vs GPT-6 Sol vs GPT-5.6 Luna
GPT-6 Luna - about $0.05 per 1,000 tickets
$0.03 input and $0.15 output per 1M tokens on AIReiter, against OpenAI's official $0.10 and $0.50. The cheapest GPT-6 tier, and the one that makes per-record automation arithmetic work.
GPT-6 Sol - about $0.91 per 1,000 tickets
$0.60 input and $3.00 output per 1M tokens, roughly 20x Luna. Worth it when a task needs flagship reasoning; wasted when the task is extraction that Luna already gets right.
GPT-5.6 Luna - the previous generation
Official $0.20 input and $1.20 output per 1M tokens. GPT-6 Luna halves the official input price and cuts output to under half, in the same cost-efficient slot.
How to choose
Run your own worst record through both tiers in the playground above. If Luna's answer holds up, the 20x gap means Sol has to earn the difference on that specific task rather than by reputation.
GPT-6 Luna FAQ
Common questions about the online playground, pricing, and API access.
/ 01When should I choose GPT-6 Luna?
Choose Luna for high-volume, well-defined work such as summarization, extraction, and classification. Escalate to GPT-6 Sol only when a task needs flagship reasoning.
/ 02How much does GPT-6 Luna cost on AIReiter?
AIReiter charges 30% of OpenAI's official rates: $0.03 per 1M input tokens and $0.15 per 1M output tokens, versus the official $0.10 and $0.50. Cached input reads are $0.003 per 1M.
/ 03How does GPT-6 Luna differ from GPT-5.6 Luna?
GPT-6 Luna is the newer generation, at half the official input price and a fifth of the official output price of GPT-5.6 Luna, in the same cost-efficient tier.
/ 04What is the context window?
GPT-6 Luna supports roughly 1M input tokens and up to 128K output tokens. Requests above 272K input tokens are billed at 2x input and 1.5x output, matching OpenAI's own policy.
/ 05How fast is GPT-6 Luna?
On AIReiter, a summarize-and-extract request over a 494-token support ticket returned 204 output tokens in 10.6 seconds without streaming, and 7.2 to 7.4 seconds end to end with streaming. Note that Luna reasons before it answers: the first streamed token took 5.0 to 5.6 seconds across runs, so most of the wait happens before any text appears. Lower the reasoning effort if you need the response to start sooner.
/ 06What does GPT-6 Luna cost per 1,000 records?
Using that same measured request - 494 input and 204 output tokens - about $0.05 per 1,000 tickets at AIReiter rates. The same workload on GPT-6 Sol costs roughly $0.91 per 1,000.
/ 07Can I call GPT-6 Luna through an API?
Yes. Follow the linked API documentation and use model ID gpt-6-luna against the OpenAI-compatible endpoint.