Frontend · Artificial Intelligence · ChatGPT

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

By Canace · Aug 6, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

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.

Summary

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.

Takeaways

— Luna is the fastest, cheapest tier for chat, translation, and simple code — comparable to Claude Haiku.

— Terra is the balanced general-purpose tier for writing, daily development, and office work — comparable to Claude Sonnet.

— Sol is the flagship tier for complex reasoning, agents, research, and architecture — comparable to Claude Opus.

— Thinking Effort is a separate parameter from model tier; it controls how long the model spends reasoning, not which model runs.

— Ultra is the highest Thinking Effort setting and can consume $2–2.50 per minute.

— Sol Ultra can autonomously decompose vague requirements into sub-tasks and dispatch them to sub-agents.

— Long-context stability improves under Sol Ultra, with fewer hallucinations during extended monorepo tasks.

— The Codex app becomes laggy when Ultra is enabled.

— Codex now attempts to self-diagnose and fix bugs, but the repair process burns the user's tokens.

— A 20× / 5-hour quota lasts roughly three hours of continuous Sol Ultra usage.

— Tibo recommends Sol Medium as the daily driver, reserving Extra High and Ultra for genuinely hard problems.

— Early reports suggest Luna's real-world performance already matches Terra's while costing less.

Conclusions

Decoupling model capability from reasoning budget is a pricing and workflow shift: it lets a weaker model think longer or a strong model answer fast, but it also offloads cost management onto the user.

The self-repair feature is a double-edged move — it reduces manual intervention but charges the user for the AI's own debugging, which feels like paying for a vendor's QA cycle.

Ultra's $2.50/min burn rate makes it a specialty tool, not a daily driver; the gap between Medium and Ultra is wide enough that most work won't justify the spend.

Luna reportedly catching up to Terra in real-world quality suggests OpenAI's tier labels may already be lagging behind actual model performance.

Concepts & terms

Thinking Effort

A parameter in GPT-5.6 that controls how much compute time the model allocates to reasoning about a problem, independent of which model tier (Sol, Terra, Luna) is selected. Settings include Medium, High, Extra High, and Ultra.

Sol / Terra / Luna

OpenAI's three fixed capability tiers for GPT-5.6 models, replacing the old Instant, Thinking, and Pro naming. Sol is the flagship, Terra is balanced, and Luna is the fast/cheap tier.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗