GPT-5.6 Luna API Guide (2026): Pricing, Context, and the 272K Fix

On July 18, 2026, OpenAI pushed a small but important Codex CLI update: version 0.144.6 refreshed the bundled instructions for GPT-5.6 Sol, Terra, and Luna and corrected their context windows to 272,000 tokens in the CLI metadata. That sounds cosmetic until you run agents at scale. Bad context metadata makes routers, truncation logic, and budget guards lie to you.

At the same time, the API docs still show GPT-5.6 Luna as a cheap, high-volume model with a much larger model context window. That mismatch is exactly why this matters. If your tooling reads the wrong number, you may over-truncate, under-budget, or send prompts that look safe in staging and explode in production.

TL;DR / Key Takeaways

If you build coding agents, support bots, or routing layers, the real lesson is simple: stop treating model names as static labels. Treat them as moving parts with pricing, context, and tool behavior that can change underneath you. A gateway like KissAPI helps here because you can swap tiers without rewriting every client.

What the July 18 Fix Actually Changes

The obvious reading is that OpenAI fixed a docs bug. The practical reading is better: your automation should not trust old cached metadata. If your router believed Luna had more or less context than it really does, it may have been making the wrong decision about chunking, summarization, or fallback selection.

That matters most in three places:

• Agent compaction: if your state machine decides when to summarize based on a stale window, you can drop important context too early.
• Prompt budgeting: if you size a request assuming the wrong maximum, your safety margin disappears fast.
• Model selection: if the cheapest model is also the one your tooling understates or overstates, your routing logic will drift.

Pricing Table

Those numbers make the playbook pretty clear. Start with Luna for the noisy, high-volume stuff. Use Terra when the task needs more headroom. Keep Sol for the jobs where quality matters more than cost. That is the cleanest way to avoid paying flagship prices for grunt work.

Model Comparison

How I Would Route This in Production

My default rule is boring, and that is why it works:

1. Send short, repetitive, or easy-to-check tasks to Luna.
2. Escalate to Terra if the task spans more context or the first pass looks shaky.
3. Escalate to Sol only when quality matters more than the extra cost.

This is where many teams overthink it. They spend two weeks benchmarking the best model and then wire every task to that model. That is a nice way to burn budget. Route by job shape, not by ego.

curl Example

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-luna",
    "input": [
      {
        "role": "system",
        "content": "You are a concise support triage assistant. Classify the request, extract the issue, and return JSON."
      },
      {
        "role": "user",
        "content": "Classify this ticket: The webhook retries every 30 seconds after a 429."
      }
    ]
  }'

Python Example

from openai import OpenAI

client = OpenAI()

resp = client.responses.create(
    model="gpt-5.6-luna",
    input=[
        {"role": "system", "content": "You are a concise support triage assistant. Return JSON only."},
        {"role": "user", "content": "Summarize this incident in one paragraph and list next steps."}
    ]
)

print(resp.output_text)

Node.js Example

import OpenAI from "openai";

const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });

const response = await client.responses.create({
  model: "gpt-5.6-luna",
  input: [
    { role: "system", content: "Return valid JSON with fields: summary, severity, action." },
    { role: "user", content: "A batch job failed because the token budget was exceeded." }
  ]
});

console.log(response.output_text);

What Developers Should Check Before Shipping

If you are already using GPT-5.6 Luna, do three things today:

• Refresh your model metadata from the live docs, not from a hardcoded enum.
• Log the effective context window your router thinks it is using.
• Put a budget ceiling on output tokens before your first production spike.

The best teams also keep a fallback path. I like a unified gateway for that because it lets you switch vendors or tiers without touching every integration. If you already use KissAPI, this is exactly the kind of swap it makes painless.

FAQ

Is GPT-5.6 Luna the cheapest GPT-5.6 model?

Yes. OpenAI's pricing page lists Luna at $1.00 per 1 million input tokens and $6.00 per 1 million output tokens.

Why does Codex CLI show a 272,000-token context window?

OpenAI's Codex CLI 0.144.6 refresh on July 18, 2026 corrected the bundled metadata for GPT-5.6 Sol, Terra, and Luna to 272,000 tokens inside the CLI. That is separate from the live API model docs.

When should I use Terra or Sol instead?

Use Terra when Luna is too weak for a task but you still care about cost. Use Sol when the task is expensive to get wrong and needs the strongest tier in the family.