Thoughts on new OpenAI models

On July 9th, 2026, OpenAI announced three new models from the GPT-5.6 family: Sol, Terra, and Luna. On one hand, I am very eager to try everything new. On the other hand, OpenAI Codex is one of my main work instruments right now, so I'm very careful with switching. For work, I use both OpenAI Codex and Claude Code, and they work well for me.

Each model is different. I'm used to GPT-5.5 in High and Extra High settings. This is my favorite. I also tried Claude Fable 5, and I work quite a lot with Opus 4.8, but I will write separately about Claude models.

Very important for me in models is the following: 

• Can they understand what actually needs to be done without inventing unrelated things themselves?
  This is a pretty interesting topic because sometimes you tell the model what they want and you ask it to write a plan. If you look at the plan, you will see that it added two or three times more than what it thinks is needed, but you never asked for it. And it is not really what you want.
• How fast do they work?
  Most of the models make work done, but not all of them do it fast enough. Some are faster, some are slower or much slower. This is an interesting trade-off with the next point. 
• What about quality? 
  Some models write the code more complicated than it has to be. Others can make trivial mistakes. Some of them go and check the documentation without you asking them to. Others do not. If you want to get a good result, you really have to watch what AI is doing. 
• Conversation style 
  I put it as the last point, but this is actually very important for me. I'm very sensitive to how people and models speak and what they say. I'm getting used to a certain style of a certain model, and switching to a new model is usually very complex for me. It just feels all wrong when you expect one style of answering and you get another. 

GPT-5.5 at high and extra high settings is my favorite for coding.

This model thinks very well. It asks questions only if it has to. It can see what you missed, and then it will tell you about it. It can also check sources on its own, including documentation for new TYPO3 versions, and it can inspect source code without you especially asking it to. This is what many other models like it do: they go and check and verify their knowledge and assumptions against the actual code.

It generates very optimized code in most cases. Very rarely it gets carried away and does something strange, but if you tell it, then it will go and fix it very quickly. 

Conversational style of this model is also really good. It's always polite, it's always very attentive, and it feels warm. It can joke, but it tries to minimize this and stay more on being warm but professional.

It is also a great reviewer of its own code as well as code generated by other models or written by humans.

I tried all three new models, and this is what I can say about them. 

Luna can do the work, but you have to watch it very carefully. This model is very cost-effective, but it comes at the price of the quality of its reasoning. If you give it a task where it needs to think and make decisions, most likely these decisions will not be good enough. This is not a thinker. This is something that can change 100 strings for you from one type of quotes to another type of quotes. That's it. It can code, but you had better not try it unless it is something very trivial. And you shouldn't use anything less than “Extra high” setting. 

Terra is a balanced general all-purpose model which you can use for normal work. I found that its code is not very different from what 5.5 did. The only thing which was different for me was the conversation style. I would say that 5.5 was warmer and more pleasant to work with. Terra is drier, and it sounds more focused. This is more personal than result-related, but if you appreciate warmer style, then you will like 5.5 more. Otherwise, I didn't notice changes between Terra and 5.5 at “High” settings. 

Sol is the largest one. I think OpenAI made it quite well with names this time. Luna is small, Terra is normal, and Sol is huge.

Sol is really powerful. However, it does not feel to be like Anthropic Fable. When you talk to Fable, you actually feel that it is big and huge and very powerful. With Sol, it is different. You really feel that it is more powerful than other models, but it keeps that warmth that OpenAI models usually have.

I made some changes to my small internal projects with Sol, and I was really satisfied with it. It works much faster than Fable, and it never made a single mistake in the code. It was very clever to discover what I missed, and it asked me very clever questions and proposed very good solutions. On one hand, it caught itself on a possibly destructive action. The action was approved by a classifier model in auto mode, but Sol still caught itself. This was very interesting. I could still recover if it removed the file or corrupted the file because I had it in git, but I never saw the model catching itself doing something dangerous. 

For the code quality, it did an amazing job. I looked at it, and I really liked it. It wrote very short code which did exactly what was needed, and it did it beautifully. It looked like a very senior programmer. Also, it wrote amazing documentation about the feature.

Today it also audited my website and made a security report about it, finding one medium-severity issue and some small issues. None of them could break the site, but they were still good to fix. Thanks to Sol.

To me, Sol feels like GPT-5.5 on steroids. 

I would probably stay with 5.5 on high or extra-high reasoning for now, and occasionally I will use Sol for more complex tasks or for reviews. I still like the style of 5.5 more than new models, and since this has a very serious meaning for me, I think I will stay on 5.5 for now. It does the work. It does it well enough, so I don't see a reason to change unless I have to make reviews or really complex investigations or really complex thinking. 

P.S. Both images on this page are generated by the Terra after the discussion about the visual representation of their power for about 20 minutes. It told me a lot of disclaimers that images do not really represent their power, but I said: “Okay, just make me nice images for my blog!” Initially I wanted to post only one image, but I couldn't decide which one, so I've added both.