GPT-5.6 Luna Guide: Nano Pricing With a Six-Step Effort Dial

The cheapest model in a family is usually judged on whether it is good enough. GPT-5.6 Luna needs a different question, because its price is not fixed in practice. The per-token rate is low, but the same model at its highest reasoning effort can produce many times the output tokens it does at its lowest. Luna is cheap when you configure it to be, and not otherwise.

OpenAI released the GPT-5.6 family on 9 July 2026: Sol for frontier capability, Terra for a balance of intelligence and cost, and Luna "for efficient, high-volume workloads". Luna's model page says it roughly corresponds to the nano tier of earlier GPT-5 families.

Where Luna sits in the GPT-5.6 line

The three tiers share most of their specification. All three list a 1,050,000-token context window with up to 922,000 input tokens and 128,000 output tokens, a knowledge cutoff of 16 February 2026, text and image input, and text output. The differences are capability and price.

• Sol at $4 input and $20 output per million tokens — the flagship, roughly the unsuffixed tier.
• Terra at $2 and $12 — roughly the old mini tier.
• Luna at $0.20 and $1.20, with cached input at $0.02.

Those are OpenAI's list prices at the time of writing, and OpenRouter shows the same Luna rate. They are not launch prices: on 30 July OpenAI cut Luna's price by 80%, so anything you read from July quotes a much higher figure. If the mid tier is closer to what you need, the GPT-5.6 Terra guide covers it, and the GPT-5.6 Sol guide covers the top.

Nano pricing with a frontier-sized window

OpenAI's own comparison table puts Luna next to GPT-5.4 nano: $0.20 input and $0.02 cached for both, $1.20 output for Luna against $1.25 for nano. On the rate card, Luna is a like-for-like replacement for the older nano.

The difference is what comes with it. Luna supports the same Responses tool list as Sol — web and file search, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search — plus Chat Completions and the Batch API. Fine-tuning is not supported. A nano-priced model with computer use and a 1M window is a different proposition from a classifier.

The window comes with the usual cost rule: prompts over 272K input tokens bill at twice the input rate and one and a half times the output rate for the whole request. Large windows are useful, but a big context is not the same as a context the model attends to well. Context length vs effective context covers why.

Effort is the real price dial

Luna accepts six reasoning effort levels: none, low, medium (the default), high, xhigh and max. That range is the main thing to understand about the model.

Artificial Analysis benchmarked Luna at max effort. At the time of writing its page records about 150 million output tokens to run its Intelligence Index, against about 90 million for GPT-5.6 Sol at the same setting, and a time to first token around 120 seconds. Even so, it put Luna's cost per index task at $0.18, against $1.99 for Sol at max. Two lessons follow. At maximum effort Luna is far more verbose and very slow. It is still roughly a tenth of Sol's cost per task.

For high-volume work you rarely want max. Classification, extraction, routing and short rewrites belong at none or low, where latency is low and output is mostly the answer. Note the default is medium, so a request that omits the field is already spending reasoning tokens. The token cost of reasoning models covers how to measure what each setting actually costs you.

A sensible starting configuration:

{
  "model": "gpt-5.6-luna",
  "reasoning": { "effort": "low" },
  "max_output_tokens": 2000
}

Raise effort per task type only when your evaluation shows the gain.

Persisted reasoning across the family

A GPT-5.6 feature that matters most for the cheap tier: OpenAI's reasoning guide says Sol, Terra and Luna can reuse each other's persisted reasoning, and GPT-5.6 models keep reasoning across turns by default. That makes mixed-tier routing more practical. A conversation can run on Luna and escalate a hard turn to Sol without throwing away the reasoning state. Reasoning does not carry over to or from the GPT-5.5 family.

That is the shape where Luna earns its keep: the default model in a router, with a stronger tier behind it. Model routing and fallbacks covers how to build the escalation rule.

GPT-6 Luna undercuts it

On 22 September OpenAI released GPT-6 Luna at $0.10 input, $0.01 cached and $0.50 output per million tokens — half Luna's input rate and well under half its output rate. Artificial Analysis now marks its GPT-5.6 Luna page as deprecated and points readers to the newer model.

So for new high-volume work, evaluate GPT-6 Luna first. GPT-5.6 Luna remains the right choice where you have a validated integration, where you want to share persisted reasoning with Sol or Terra, or where your own tests favour it. The GPT-6 Luna guide covers the successor.

Variants and the Pro setting

OpenRouter lists a :batch variant at $0.10 input and $0.60 output at the time of writing: the same model on a discounted asynchronous queue, for work that can wait. It also lists GPT-5.6 Luna Pro, which is not a different model but Luna with reasoning.mode set to pro, billed at the same token rates. The GPT-5.6 Luna Pro guide covers when that setting is worth its extra tokens; for most high-volume jobs, plain Luna at low effort is the right default.