GPT-6 Luna Guide: OpenAI's Cheapest GPT-6 Tier in Practice

Most teams pick their default model by looking at the top of a leaderboard and working down until the price stops hurting. That approach skips the tier that handles most real traffic: classification, extraction, routing, summarising, short edits. GPT-6 Luna is OpenAI's model for that tier, and its price is low enough that the interesting questions are about configuration rather than cost.

OpenAI released Luna alongside GPT-6 Sol on 22 September 2026, a few weeks after the flagship GPT-6 Astra. Its model page calls it the company's most efficient model for focused, high-volume tasks. That is an accurate description of the job, and a useful warning about what it is not for.

Where Luna sits in the GPT-6 line

The GPT-6 family has three tiers. Astra is the flagship. Sol is positioned as the cost-efficient high-end model for complex coding and agentic work. Luna sits below both. OpenAI says Sol and Luna were trained with methods similar to Astra's and that both improve on the equivalent GPT-5.6 models in professional work, factuality, coding and computer use, while using fewer tokens.

The published specification is more generous than the tier suggests:

• Context window: 1,050,000 tokens, with a 922,000-token input limit.
• Maximum output: 128,000 tokens.
• Modalities: text and image in, text out.
• Knowledge cutoff: 18 May 2026, slightly later than Sol and Astra.
• Reasoning effort: none, low, medium (default), high, xhigh and max.

The context window and output limit are the same as Astra's. What separates the tiers is how well they reason over that context, not how much of it they accept.

The price, and what it buys at volume

OpenAI's list price is $0.10 per million input tokens and $0.50 per million output, with cached input at $0.01. OpenRouter lists the same figures at the time of writing. The predecessor, GPT-5.6 Luna, was reported at $0.20 input and $1.20 output, so input halved and output fell by more than half.

To make that concrete: a ticket-triage job that reads 2,000 tokens and writes 150 costs roughly $0.000275 per ticket at list price. A million tickets is about $275. At that scale, the variable that moves your bill is not the rate. It is whether you are sending the same system prompt uncached a million times, which is where the $0.01 cached-input rate earns its keep.

The catalogue also lists openai/gpt-6-luna:batch, the same model on an asynchronous endpoint at half price: $0.05 input and $0.25 output per million tokens. OpenAI applies the same 50% discount to its own Batch and Flex processing. For overnight enrichment or back-filling a dataset, there is little reason to pay the synchronous rate.

Reasoning effort is the dial that matters

Luna defaults to medium effort. For a large share of high-volume work that is more reasoning than the task needs. Classifying a support email or pulling fields from an invoice rarely benefits from the model deliberating first, and every reasoning token is billed as output.

Start at none or low for extraction and labelling, measure accuracy, and only raise effort where errors appear. OpenAI's own launch claim runs the other direction and is worth knowing: it says Luna at max effort beats GPT-5.6 Sol at medium for roughly a tenth of the cost. Treat that as a vendor claim, but it suggests Luna at high effort is a credible substitute for last generation's mid-tier on bounded reasoning tasks.

One API detail catches people. On OpenAI's Chat Completions endpoint, Luna supports function calling only when reasoning effort is none. If you need tools and reasoning together, use the Responses API, which also exposes the hosted tools such as web search, file search and code interpreter.

A million tokens, priced in two tiers

A 1M window on a model this cheap is tempting for whole-repository or whole-archive prompts. Read the pricing note first. Prompts over 272K input tokens are billed at twice the input and cache rates and one and a half times the output rate, for the entire request, not only the tokens above the threshold.

That still leaves Luna cheap in absolute terms, but it changes the comparison with retrieval. A 600K-token prompt sent repeatedly is often worse value than retrieving the 30K tokens that matter, and a large window does not guarantee the model attends well to all of it. The token cost of long context covers when filling the window is worth it.

Luna or Sol: a decision rule

Sol lists at $2 input and $10 output, twenty times Luna on both. That gap is large enough that the default should be Luna, with escalation where it fails.

• Pick Luna for classification, extraction, routing, summarising, short rewrites, first-pass triage and anything you run millions of times.
• Pick GPT-6 Sol for multi-file coding, long agent loops and tasks where a wrong answer costs a human more than a minute to notice and fix.
• Try GPT-6 Luna Pro first when Luna is close but not quite reliable enough. It keeps Luna's rates and spends more tokens per request, which can be cheaper than moving up a tier.

A first-week evaluation plan

Pull two hundred real inputs from the job you want to move. Run them through Luna at none, low and medium, and through whatever model you use today. Score accuracy against a labelled sample, and record output tokens per request so you can see what each effort level actually costs.

If Luna at low effort matches your current model, you have probably just cut that job's bill by an order of magnitude. If it does not, route the failures to a stronger model rather than abandoning Luna for the whole workload. Model routing and fallbacks covers how to set that up without much code.