GPT-5.6's Sol, Terra, and Luna Models, and When Ultra Thinking Is Worth the Cost
That month Codex updated, not only changing its logo but also its model naming convention. The previous Instant, Thinking, and Pro have now become Sol, Terra, and Luna, plus Medium, High, and Ultra, which is confusing.
I looked up some information and combined it with my own experience to try to explain these concepts clearly, hoping that after reading this, you'll know what they represent and how to choose.
Starting with GPT-5.6, OpenAI gradually changed the original Instant / Thinking / Pro into three fixed names:
• Sol (Sun) • Terra (Earth) • Luna (Moon)
Officially, numbers represent the model generation (5.6), and names represent the model capability tier. This way, future upgrades only need to go from GPT-5.6 to GPT-5.7, without inventing a new naming scheme each time.
Personally, I find this naming convention easy to remember. When switching models, I think: Who is the biggest in the universe? Luna → Terra → Sol, and I know the general capability level.
Let's look at their positioning separately.
1. What are the differences between Sol, Terra, and Luna?
• Luna: Fastest and cheapest, suitable for chat, translation, simple code, and batch tasks. Comparable to Claude's Haiku. • Terra: The most balanced overall capability, suitable for writing, daily development, office work, and most scenarios. Comparable to Claude's Sonnet. • Sol: The current flagship model of the GPT-5.6 family, suitable for complex reasoning, agents, scientific research, architectural design, and long-code high-difficulty tasks. Comparable to Claude's Opus.
2. What is Thinking Effort?
Besides the models themselves, there's another concept that many people easily confuse: Thinking Effort.
It is not a new model, but a parameter that controls how much time the model is willing to spend thinking.
A simple understanding:
• Sol / Terra / Luna determines the model's inherent capability. • Thinking Effort determines how much thinking time the model is willing to invest.
So the two are actually different dimensions.
The newly added Ultra is currently the highest tier of Thinking Effort.
Just seeing the name feels domineering, so I had to try it.
So I ran Sol + Ultra for a few days and will share a few things that impressed me.
3. What does Sol Ultra feel like to use?
1. The agent starts to proactively work
Vague requirements can be automatically broken down into tasks and distributed to multiple different subagents for execution; I only need to handle the acceptance.
It truly frees up your hands, letting you sip tea while the work gets done.
Previously, I had to write the requirements document myself, break down the tasks, and semi-automate the agent to do the work. This really improves efficiency significantly.
2. Long-context capability is noticeably more stable
Another thing I'm quite satisfied with is the performance on long tasks.
For example, in a relatively complex Monorepo project, a task ran continuously for a long time, and Sol Ultra's understanding of the context remained relatively stable.
This point is actually very important in engineering scenarios.
In the past, when tasks ran for a long time, they would often start to "go off track" or produce inexplicable hallucinations.
This time, I clearly felt that this situation was much less frequent.
Codex lead Tibo also mentioned on X that he hopes people can spend more time on truly ambitious tasks, rather than constantly coming back to correct the agent. My own experience is quite consistent with this statement.
4. Does Ultra have any drawbacks?
Of course, and quite a few.
First is performance.
Many people report that after enabling Ultra, the Codex App becomes laggy.
Interestingly, though, Codex has now started trying to "fix itself."
It has added capabilities similar to OpenClaw and Hermes Doctor, where after discovering a problem, it will try to locate and fix the bug itself.
Of course, these repair processes still consume your tokens. It always feels like spending your own money to treat the AI's illness, which is a bit impolite of OpenAI.
The second issue is cost.
A friend tested it and found that Sol Ultra consumes about $2 to $2.5 per minute. A 20× / 5-hour quota can only run continuously for about three hours.
For people who heavily use agents, the cost is still quite high.
So Tibo's suggestion is:
Usually, use GPT-5.6 Sol Medium as your main model; switch to Extra High when you encounter a truly tricky problem; if you pursue the best results and don't mind quickly burning through your quota, then Ultra is indeed a beast.
Take the advice: if it's just daily chat, coding, or writing articles, the default configuration of Terra or Sol is sufficient.
When you truly encounter complex projects, long-running agents, or need to maximize quality in one go, then enabling Ultra will be more cost-effective.
This article was written shortly after the release. I saw a few days ago that Luna's actual tested experience has already reached Terra's level, and it's cheaper, so you can give it a try.