The Vibe Father
First look
Every three-tier model family has a middle child, and the middle child is usually the one you actually run all day. In the GPT-5.6 lineup OpenAI shipped on July 9, 2026, that's Terra: the mid-tier, priced between the flagship Sol and the budget Luna, and — if the pattern of every recent lab holds — the model most teams will spend the most money on without ever putting on a slide.
We run a live coding leaderboard, so we care less about which model wins a headline and more about which one wins your monthly bill. Terra is squarely in that conversation. Here is what is confirmed, what is not, and where we would seat it before the independent numbers arrive.
What OpenAI confirmed about Terra
Terra is the mid tier of the family, priced at $2.50 per million input tokens and $15 per million output — exactly half of Sol's $5/$30 on both sides. That symmetry is not an accident; it is OpenAI telling you where Terra sits before it tells you anything about how it performs.
Like the rest of the 5.6 family, Terra is built for bigger multi-step, long-horizon agentic work with less hand-holding, and it inherits the new max reasoning-effort setting. OpenAI's marquee coding claims — the "best coding model yet" line, the Terminal-Bench 2.1 state-of-the-art number — are all attached to Sol, not Terra. That is important. OpenAI has told us where Terra is priced and roughly what it is for; it has not published a coding benchmark for it.
What is still unknown
Terra's actual coding numbers. There is no independent — or even OpenAI-reported — SWE-bench Verified, Terminal-Bench, or LiveCodeBench figure for Terra at time of writing. So Terra does not have a score on our board, and we will not invent 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%.
Why does the mid tier so often ship without its own benchmark chart? Because labs benchmark on what flatters the family, and the flagship flatters best. The mid tier's real story is a ratio — performance retained versus price paid — and that ratio only becomes visible once independent evals score all three tiers on the same tests. Until then, Terra's value is a hypothesis.
Where Terra likely belongs on an agent team
The mid tier's classic seat is the volume builder — the model that does the bulk of the actual implementation once a smarter, pricier model has drawn the plan. Sol thinks; Terra types. On a team that runs a planner-plus-builders pattern, Terra is the workhorse that turns an approved plan into diffs across dozens of turns, where its half-price output rate compounds into real savings versus running the flagship for the same grind.
At $15 output, Terra is cheap enough to keep busy but not so cheap you would hand it your entire pipeline the way you might a scout model. Think of it as the reliable mid-level engineer on the team: not the architect, not the intern, the one who ships the majority of the code. We map every seat in best model for each agent role, and the mid-tier builder chair is one of the most cost-sensitive on the board.
How to evaluate Terra yourself the day it lands
You do not need our board — or OpenAI's — to know whether Terra is right for your volume work. The mid tier is the one where a self-run evaluation pays off most, because its whole pitch is a ratio only your workload can measure. The loop:
- Pick one real, representative task. Not the hardest thing in your backlog — a typical one. Terra's job is the median ticket, so evaluate it on the median ticket: a normal feature with acceptance criteria, a routine refactor behind a test suite.
- Drop Terra in behind a verification gate. Same harness, same context, same tools. Change only the model. The point is to isolate Terra's contribution, not to run a new setup.
- Let your tests judge. Build green? Suite green? Did it stay inside the files it was asked to touch? A builder that survives verification is worth more than one that produces prettier prose about what it did.
- Compare cost against your current builder on the same task. This is the decisive step for a mid tier. Measure end-to-end task cost — tokens, retries, review time — for Terra versus whatever does volume implementation today. Half the flagship's rate only matters if Terra needs a comparable number of tokens to finish; a cheaper model that loops forever is not cheaper.
Run that on a handful of representative tickets and you will know Terra's true performance-per-dollar on your code long before a consensus forms online.
Day-one access without switching tools
The reason we can test Terra the hour its API opens is that testing a new model should be a dropdown, not a project. The Vibe Father is a model-agnostic macOS command deck running 22 CLIs side by side, so Terra is a drop-in the moment OpenAI ships access — point an agent at it on a branch, run the loop above, keep your current builder doing production work in the meantime. If Terra wins the volume seat on cost, you promote it in minutes. If Luna or an existing model beats it per dollar, you have lost an afternoon, not a migration.
Our read on Terra today: the tier most likely to matter to your bill and least likely to make your feed, which is exactly backwards from how attention gets allocated on launch day. Sol will win the arguments this week; Terra may quietly win the budget this quarter — but only if the numbers justify it, and those numbers do not exist yet. When they do, they will be on the leaderboard beside its siblings, and the honest comparison to make is Terra-versus-mid-tier-rivals, not Terra-versus-Sol. For the full setup decision — which CLI to run Terra through in the first place — start with how to choose an AI coding CLI.
