GPT-5.6 Luna, Explained

GPT-5.6 Luna is the fastest and most affordable tier in OpenAI's GPT-5.6 family. On July 30, 2026, OpenAI cut its price by about 80%, to $0.20 per million input tokens and $1.20 per million output tokens.

Luna is built for high-volume, everyday tasks: summarizing text, drafting first copy, sorting and tagging content, and running routine automation. It is not the tier for the hardest reasoning — that is what Terra and Sol are for.

This guide explains what Luna is, what it is good at, its price-performance story, when to pick it over Terra or Sol, and the workloads its new low price finally makes affordable.

What is GPT-5.6 Luna?

GPT-5.6 Luna is the entry tier of OpenAI's GPT-5.6 model family, tuned for speed and low cost. It sits below Terra (balanced) and Sol (flagship) in the same lineup.

OpenAI announced GPT-5.6 on June 26, 2026, and made the family generally available in the API and Codex in July 2026. There is no waitlist and no vetting to use Luna.

Luna trades some deep-reasoning power for much faster responses and a far lower per-token price. For simple, repeatable jobs, that trade is usually worth it.

• Tier role: fastest and cheapest of the three GPT-5.6 tiers.
• Availability: generally available in the OpenAI API and Codex since July 2026.
• Access: no waitlist, no approval process.
• Siblings: Terra (balanced, high-volume business work) and Sol (flagship reasoning).

What GPT-5.6 Luna is built for

Luna is built for high-volume, low-complexity work where speed and cost matter most. These are jobs you run thousands of times, not one hard problem you think through once.

Its sweet spot is text that needs shaping, sorting, or shortening rather than novel reasoning. On those tasks, Luna returns answers fast and cheap enough to run at scale.

If a task has a clear right answer and a simple shape, Luna usually handles it well. If it needs multi-step planning or judgment, move up to Terra or Sol.

• Summarization — condensing articles, transcripts, tickets, and long threads.
• Drafting — first-pass emails, product copy, replies, and outlines.
• Classification — tagging, routing, and sorting content by topic or intent.
• Routine automation — repeatable steps in a workflow with a known pattern.
• Cheap high-volume agents — background agents that run constantly at low cost.

The price-performance story

Luna delivers performance comparable to models that were frontier-class a year ago at roughly 6 cents on the dollar per task, at nearly 9 times the speed. That is OpenAI's own framing from its July 30, 2026 price-performance post.

On Agents' Last Exam, OpenAI reports that Luna outperforms Claude Fable 5 at an estimated cost per task nearly 99% lower. That combination — comparable or better results at a tiny fraction of the cost — is the whole pitch.

The July 30 cut dropped Luna's input price from $1.00 to $0.20 and its output price from $6.00 to $1.20 per million tokens. That is an 80% reduction, and it changes the math for many workloads.

• Input: $1.00 → $0.20 per million tokens (about 80% lower).
• Output: $6.00 → $1.20 per million tokens (about 80% lower).
• Speed: roughly 9× faster than the year-ago frontier class it matches.
• Cost per task: roughly 6 cents on the dollar vs that year-ago frontier class.
• Agents' Last Exam: beats Claude Fable 5 at an estimated ~99% lower cost per task.

Luna vs the cheap-tier rivals

Luna competes with the budget tiers from other labs — mainly Claude Haiku 4.5 and Gemini Flash. All three target the same job: fast, cheap, high-volume text work.

OpenAI's price cut is widely read as a response to exactly this pressure, from Gemini's Flash tier, Claude Haiku, and cheap open-weight models. At $0.20 input and $1.20 output, Luna is priced to win on cost.

Pick the tier that fits your stack and your quality bar. If you already run on one cloud or platform, the native cheap tier there may be simpler even if list prices differ. Test on your own tasks before you commit.

• Luna vs Claude Haiku 4.5 — two budget tiers aimed at the same high-volume jobs; see the full breakdown.
• Luna vs Gemini Flash — another fast, cheap rival; compare speed, price, and fit.
• General rule: benchmark all three on your real prompts, not on marketing numbers.

When to pick Luna over Terra or Sol

Pick Luna when the task is simple and the volume is high. Its speed and low price pay off most when you run the same kind of job again and again.

Choose Terra when tasks need more judgment — customer support, internal tools, and document analysis that Luna handles unevenly. Terra is the balanced middle tier for everyday business work.

Choose Sol for the hardest work: complex coding, cybersecurity, science, and deep multi-step reasoning. Sol adds new max and ultra reasoning modes, and ultra can spawn subagents.

• Luna — high-volume summarizing, drafting, classifying, and routine automation.
• Terra — balanced everyday business work: support, internal tools, doc analysis.
• Sol — flagship reasoning: complex coding, cybersecurity, science, deep planning.

When Luna is not enough

Luna is not enough when a task needs deep reasoning, careful judgment, or long multi-step planning. On those jobs, its speed-and-cost trade starts to hurt output quality.

The warning signs are easy to spot: answers that miss nuance, plans that skip steps, or code that looks right but breaks. When you see them, escalate to Terra, and to Sol for the hardest cases.

A common pattern is to route by difficulty. Let Luna handle the easy majority, and pass the hard minority up to Terra or Sol automatically. That keeps costs low without capping quality.

• Escalate to Terra — when tasks need real judgment or context that Luna keeps missing.
• Escalate to Sol — for complex coding, security, science, or deep reasoning.
• Route by difficulty — Luna for the easy majority, higher tiers for the hard tail.

Workloads the price cut unlocks

The 80% cut makes several high-volume workloads affordable that were borderline before. When each task costs a fraction of a cent, you can run jobs at a scale that used to blow the budget.

The biggest change is always-on and per-item automation. Summarizing every ticket, tagging every document, or running a background agent on every record becomes cheap enough to leave running.

Before you scale, put guardrails in place: sample the output for quality, cap spend, and route hard cases up a tier. Cheap tokens are only a win if the results still hold up.

• Summarize every support ticket or call transcript, not just a sample.
• Tag and route large document sets item by item, continuously.
• Run background agents that watch and process records around the clock.
• Add cheap drafting to every workflow step instead of a few high-value ones.
• Enrich or clean large datasets row by row at a price that pencils out.