Model guide

GPT-5.6 Luna: What the Budget Tier Is Actually Good For

Luna is the weakest GPT-5.6 tier on agentic benchmarks, and OpenAI does not pretend otherwise. But at $1 in / $6 out it changes the economics of a whole class of work. Here is where it earns its keep, and where it quietly costs you more.

By Kylian Migot · Updated July 2026 · 6 min read

Quick answer

GPT-5.6 Luna is the budget/fast tier of OpenAI's GPT-5.6 family: $1 in / $6 out per MTok, which is 2.5x cheaper than Terra and 5x cheaper than Sol. It scores 84.7% on Terminal-Bench 2.1, the weakest of the three tiers, so use it for mechanical, high-volume work, and never for tasks hard enough to fail.
What it is
Budget/fast tier of the GPT-5.6 family (Sol › Terra › Luna)
Price
$1 in / $6 out per MTok
Headline score
84.7% on Terminal-Bench 2.1 (Terra: 87.4%, Sol: 88.8%)
Not included
Max reasoning effort and ultra mode, Sol only
01

What Is GPT-5.6 Luna?

GPT-5.6 Luna went generally available on July 9, 2026, alongside Sol and Terra. In OpenAI's naming scheme the number is the generation and the name is the tier; Luna is the bottom of the three, the budget/fast option. The API identifier is gpt-5.6-luna, and it is selectable in Codex like the rest of the current generation.

The pitch is not capability, it is arithmetic: $1 in / $6 out per MTok makes Luna the cheapest way to run the current OpenAI generation, and for a specific class of work, mechanical, well-specified, high-volume, that price is the feature. The rest of this page is about drawing that class precisely, because the same price becomes a trap one step outside it. For the full lineup context, see the models hub.

02

Specs: The Verified Data

SpecGPT-5.6 Luna
VendorOpenAI
API identifiergpt-5.6-luna
Generally availableJuly 9, 2026
Price$1 in / $6 out per MTok
TierBudget / fast (bottom of the GPT-5.6 family)
Terminal-Bench 2.184.7% (OpenAI launch materials)
CodexSelectable in Codex (current generation)

Data verified July 18, 2026 against OpenAI's published pricing and launch materials. Family context: Luna trails Terra by 2.7 points on Terminal-Bench 2.1 (84.7% vs 87.4%) and, like Terra, has neither max reasoning effort nor ultra mode, those are Sol exclusives.

Ship your first agent today

Download AIDEN free and point it at your existing Claude Code or Codex setup. No credit card, running in minutes.

Download AIDEN free

Free to start · macOS 12+ · No credit card required

03

The Cost Math That Justifies Luna

Take a realistic mechanical task, a lint sweep, a rename, a formulaic edit, consuming 100k input and 20k output tokens. Per task the absolute numbers look trivial; multiplied by a pipeline they stop being trivial:

100k in / 20k outGPT-5.6 LunaGPT-5.6 TerraGPT-5.6 Sol
Price per MTok$1 in / $6 out per MTok$2.5 in / $15 out per MTok$5 in / $30 out per MTok
Cost per task$0.22$0.55$1.10
1,000 tasks$220$550$1,100

That is the whole argument: at 2.5x below Terra and 5x below Sol on both sides of the meter, a 1,000-task pipeline costs $220 on Luna versus $550 on Terra and $1,100 on Sol. When the tasks are easy enough that all three tiers would succeed, paying the Terra or Sol rate buys nothing. Budgeting across a whole Codex plan is covered in the best Codex model guide.

04

What Luna Is Actually Good For

Mechanical edits

Lint fixes, renames, import reshuffles, config tweaks: tasks you could almost script, where the spec fully determines the diff and the model just types faster than you.

High-volume small tasks

Batch work across many files or repos, where per-task cost dominates. At $1 in / $6 out per MTok, volume is exactly where Luna's pricing compounds.

Cost-capped pipelines

Automated jobs with a hard budget: triage, summarization, formulaic codemods. Low latency helps here too, Luna is the fast tier as well as the cheap one.

The common thread: the task must not need the 2.7 points Luna gives up to Terra. If success depends on multi-step reasoning, escalate before you start, to Terra for normal work, or to Sol for the genuinely hard stories. The three-way decision rule is laid out in Sol vs Terra vs Luna.

05

When Luna Is a False Economy

The practical fix is routing, not loyalty to a tier. Decide per task, and let outcomes tune the rule: mechanical cards to Luna, default work to Terra, hard stories to Sol. That is the model AIDEN operationalizes, your Codex CLI on a kanban board where each story runs its own model on its own branch, so a wrong routing costs you one card, not a pipeline.

FAQ

What is GPT-5.6 Luna?
GPT-5.6 Luna is the budget/fast tier of OpenAI's GPT-5.6 family, generally available since July 9, 2026 and selectable in Codex. It costs $1 in / $6 out per MTok, the cheapest tier in the family, and scores 84.7% on Terminal-Bench 2.1, behind Terra (87.4%) and Sol (88.8%). It is built for quick mechanical edits and high-volume cheap tasks, not hard agentic work.
How much cheaper is Luna than Terra and Sol?
2.5x cheaper than Terra and 5x cheaper than Sol, on both input and output: $1 in / $6 out per MTok versus Terra's $2.5 in / $15 out per MTok and Sol's $5 in / $30 out per MTok. On a small task consuming 100k input and 20k output tokens, that is about $0.22 on Luna, $0.55 on Terra, and $1.10 on Sol.
Is Luna good enough for serious coding work?
Honestly: it is the weakest GPT-5.6 tier on agentic benchmarks, at 84.7% on Terminal-Bench 2.1, 2.7 points behind Terra. For mechanical, well-specified tasks that is fine, the task barely exercises the gap. For hard multi-step work it is the wrong tier: a Luna failure plus a re-run on a better model costs more than starting on Terra.
When is Luna a false economy?
Whenever the task is hard enough that Luna might fail. A failed run is not free: you pay for the tokens, then pay again for the re-run, plus your own time triaging. One Luna attempt (~$0.22 on a 100k/20k task) followed by a Terra re-run (~$0.55) already costs $0.77, more than the $0.55 Terra would have cost succeeding once. Cheap only stays cheap when it succeeds.
Does Luna have ultra mode or max reasoning effort?
No. Both are exclusive to GPT-5.6 Sol ($5 in / $30 out per MTok); neither Luna nor Terra has them at any price. If a task needs the family's highest reasoning budget or ultra mode's parallel subagents, Luna is two tiers away from the right answer.

Keep reading

Send only the cheap work to the cheap tier.

AIDEN is a macOS app that orchestrates your existing Codex and Claude Code CLIs on one kanban board, BYOK, local-first, spec-gated. Free for one project.

macOS 12+ · Bring your own Claude Code or Codex · Your code stays local