GPT-5.6's Sol, Terra, and Luna Models, and When Ultra Thinking Is Worth the Cost

The split between model capability and thinking budget means developers now control cost and depth independently. Choosing wrong burns quota fast — Ultra at $2.50/min empties a subscription in an afternoon — but using it on the right problem produces agent behavior that actually finishes multi-step work unattended.

The old Instant, Thinking, and Pro labels are gone. GPT-5.6 models now use Luna (fast/cheap, Haiku-class), Terra (balanced, Sonnet-class), and Sol (flagship reasoning, Opus-class). A separate Thinking Effort parameter controls how much compute the model spends on a problem, with Ultra as the highest setting.

Running Sol with Ultra effort turns the agent proactive: it breaks vague requirements into sub-tasks and dispatches them to sub-agents. Long-context stability in monorepo projects improves noticeably, with fewer hallucinations during extended runs. Codex lead Tibo confirmed the goal is less babysitting and more ambitious work.

The trade-offs are real. Ultra makes the Codex app laggy, and the model has started trying to self-repair bugs — using your tokens. At roughly $2–2.50 per minute, a quota can drain in about three hours of continuous use. The practical advice is to stick with Sol Medium for daily work and reserve Ultra for complex, long-running agent tasks.