The Vibe Father
First look
The budget tier is where most real coding workloads actually live, which is why it is the tier we watch most carefully — and the one labs talk about least. In the GPT-5.6 family OpenAI shipped on July 9, 2026, the budget seat belongs to Luna: the light model, the cheapest of the three, and the one that might quietly punch above its price if OpenAI has done what it did with GPT-5.3 Codex.
We run a live coding leaderboard, and the budget end is where we see the biggest gap between reputation and reality. Cheap models get dismissed as toys until an independent eval shows one embarrassing the flagships on cost per solved task. So we take Luna seriously by default. Here is what is confirmed, what is not, and where it likely fits.
What OpenAI confirmed about Luna
Luna is the budget tier, priced at $1 per million input tokens and $6 per million output. That is a fifth of Sol's input rate and a fifth of its output rate — a genuinely cheap model from a frontier lab. For context on our board, that undercuts GPT-5.3 Codex ($1.75/$14) on both sides and sits in the same neighborhood as the models that currently win the price-per-solved-task fights.
Luna inherits the family's design goals — bigger multi-step agentic work with less hand-holding — and the new max reasoning-effort setting, which is more interesting on a budget model than it sounds: if you can dial a cheap model's reasoning up for the occasional hard turn, you blur the line between tiers. OpenAI's headline coding claims, though, are all pinned to Sol. Nothing about Luna's coding performance has been benchmarked publicly.
What is still unknown
Luna's actual numbers, entirely. No SWE-bench Verified, no Terminal-Bench, no LiveCodeBench figure exists for it — not independent, not even OpenAI-reported. So Luna has no score on our board, and we will not fabricate one. Independent numbers land on /benchmarks as public evals publish, scored the way we score everything: SWE-bench Verified 40%, Terminal-Bench 30%, LiveCodeBench 30%.
The "might punch up" in the title is a hope with a precedent, not a claim. GPT-5.3 Codex proved a cheap OpenAI model can post a LiveCodeBench score (87.3) that beats models many times its price. If Luna repeats that trick, it becomes one of the most important models in the family for people who ship on a budget. If it does not, it is a competent scout and nothing more. We do not know which yet, and neither does anyone posting confident charts this week.
Where Luna likely belongs on an agent team
The budget tier's natural seat is the scout and high-volume builder — the model you can afford to run constantly. Scouting means the cheap, parallel reconnaissance work: reading files to answer "where is this handled," summarizing a subsystem, drafting a first pass you expect to revise, triaging which of forty tickets is even real. High-volume building means the routine, low-ambiguity edits a plan has already de-risked. At $6 output, you can point Luna at that work all day and barely feel it.
What we would not do on day one is hand Luna the planning chair on a hard problem. The budget tier is where the SWE-bench gap to the flagships usually shows up worst — deep, cross-file reasoning is exactly what cheap models trade away. Use Luna for breadth and volume; reserve the expensive thinking for Sol. That split — scout cheap, plan expensive, build in the middle — is the backbone of every efficient agent team, and we lay it out fully in best model for each agent role.
How to evaluate Luna yourself the day it lands
Budget models reward self-evaluation more than any other tier, because their whole case is cost-per-outcome and only your workload knows your outcomes. The loop:
- Pick high-volume, representative work. Luna's audience is people running a model a lot, so test it on a lot — a batch of routine tickets, a directory's worth of scouting queries, the kind of task you would otherwise run dozens of times.
- Drop Luna in behind a verification gate. Same harness, same tools, one variable changed. A cheap model unsupervised is a liability; a cheap model behind a test gate is an asset, because the gate catches the misses cheaply.
- Let your tests judge, and count the retries. A budget model that needs three attempts to pass the suite is not a budget model. Track pass rate and retry count, not just per-token price.
- Compute cost per solved task, not cost per token. This is the only number that matters for Luna. A model at $6 output that solves it on the first try beats one at $14 that solves it on the third — and the reverse can also be true. Measure it on your work.
Do that across a representative batch and you will know whether Luna is a punch-up story or a scout, on your code, before the internet finishes arguing.
Day-one access, no tool switch
We can run that batch the hour Luna's API opens because trying a new model should cost an afternoon, not a migration. The Vibe Father is a model-agnostic macOS command deck running 22 CLIs side by side, so Luna is a drop-in the moment access ships — assign it the scout and volume seats on a branch, run the loop, and if it earns those seats on cost per solved task, promote it in minutes.
Our read on Luna today: the sleeper of the family. Nobody will lead their launch coverage with the budget model, and the budget model is where the leverage hides for teams that ship on volume. But leverage is what the numbers might show, not what they do show, because there are no numbers. When there are, they will be on the leaderboard, judged against the other cheap models that already win this fight. If you are still deciding which CLI to run Luna through, how to choose an AI coding CLI is the right first read.
