GPT-5.6 Terra: The Codex Workhorse, Explained

GPT-5.6 Terra went generally available on July 9, 2026, alongside Sol and Luna. Under OpenAI's new naming scheme the number is the generation and the name is the tier, and Terra is the middle one: below Sol, the flagship, above Luna, the budget tier. The API identifier is gpt-5.6-terra.

OpenAI's own positioning is unusually direct: Terra delivers roughly GPT-5.5 performance at 60% lower token prices. GPT-5.5 was the flagship until July 9, 2026; getting its class of capability at $2 in / $12 out per MTok instead of $5 in / $30 out per MTok is the headline of this tier, and the reason Terra, not Sol, is the sane starting point for most work. Where it sits in the full lineup lives at the models hub.

Data verified August 29, 2026 against OpenAI's published pricing and launch materials. And in family context:

Here is the arithmetic that makes Terra the default. On Terminal-Bench 2.1, the benchmark OpenAI led its launch with, Sol scores 88.8% and Terra scores 87.4%. That is a 1.4-point gap. Sol costs 2x on input and 1.67x on output: $4 vs $2 per million input tokens, $20 vs $12 on output. Paying that premium for 1.4 points only makes sense when those points are the difference between a task succeeding and failing, which for routine implementation, tests, and refactors they usually are not.

The same logic applies against the previous generation. GPT-5.5 is still selectable in Codex at $5 in / $30 out per MTok, the same price as Sol, for performance OpenAI itself says Terra roughly matches at 60% lower token prices. If you are still defaulting to GPT-5.5 out of habit, Terra is the upgrade that also cuts your bill; the full picking logic, including reasoning effort, is in the best Codex model guide, and the budgeting side in Codex usage limits.

Terra's tradeoff is a short list, but a real one: no max reasoning effort and no ultra mode. Both are exclusive to Sol. Max reasoning effort is the family's highest thinking budget; ultra mode fans a task out to parallel subagents, and OpenAI's launch materials credit it with lifting Sol's Terminal-Bench 2.1 score further still. You cannot buy either on Terra at any price.

So the escalation rule is about failure cost, not prestige. Escalate to Sol when a failed Terra run plus a re-run would cost more than one Sol run succeeding: genuinely hard cross-cutting problems, long autonomous sessions where a mid-run derailment is expensive, and terminal-heavy chains where Sol's lead is measured. For everything else, pocket the 2x input / 1.67x output difference.

Luna is 2.7 points behind Terra on Terminal-Bench 2.1, a bigger drop than Sol-to-Terra, so treat it as a tier for tasks you could almost script, not a cheaper Terra. In practice the tiers work best as a routing decision made per task, not per project: AIDEN runs your Codex CLI on a kanban board where each feature picks its own model, so the mechanical cards go to Luna, the default work to Terra, and the hard features to Sol, with each result on its own branch telling you whether the routing was right.