GPT-5.6 Sol has been around for almost a week, and I have had access to it for roughly that long. These are my honest thoughts on the model and how it has behaved in my workflow.
Day one was tough. Sol felt very different from GPT-5.5, which remains one of my favourite models. With 5.5, you normally understood what you were working towards. It could act amazingly, then five minutes later act like a complete maniac. Sol still has a little of that unpredictability, but nowhere near as much.
I still love 5.5 dearly. It was a fantastic model to work with. Sol is simply a different kind of collaborator.
It Knows What to Ask
When I proposed an idea to 5.5, it would usually accept the direction and start building. It barely asked clarifying questions.
Sol spends much more time asking questions I would never have thought to ask myself. That can initially feel slower, but those questions often expose decisions I have not made yet. Without them, the product would probably stay exactly as I first imagined it, including the parts that needed more thought.
That is one of Sol's best qualities: it knows what to ask before it starts building.
It is also much better at finishing the job. With 5.5, pressing Implement on a plan could mean getting most of the work done brilliantly, followed by a lazy final stretch where the details were quietly dropped. Sol keeps working until the plan is actually complete. That persistence is one of the main reasons I enjoy using it.
A Better Personality
I have spent most of the last month using ChatGPT. Plenty of people prefer Claude or Gemini, which is fair, but I like the way OpenAI's models communicate.
GPT-5.4 had major personality issues. It often felt as though it wanted to strike an idea down regardless of what you said. GPT-5.5 became much warmer, although it could still be overly negative. Sol is one of the first models that consistently says, in effect, "That is a good idea, but make sure you have considered this."
I actively want that from a model. Blind praise is useless, but so is automatic negativity. The best collaborator can support the direction while keeping you grounded.
I think OpenAI learned a lot from GPT-4o, which people still praise for its personality. I am not judging anyone for preferring a particular model. If you like it, you like it. For me, working with Sol has been great.
Usage Limits and Reasoning Effort
The first day had a real problem: Sol hit usage limits constantly. I am fairly efficient with my workflows and still ran into them far more than expected.
That now seems to have improved. Based on the reports I saw, the thinking budget was reduced quite significantly. Honestly, that made the model better to work with. More thinking is not automatically better thinking.
I rarely use XHigh, Max, or Ultra reasoning. Low is usually my preferred setting, with Medium as the step up when a task genuinely needs it. I tried Ultra once and it was fine, but I would rather work with one fast, responsive model than wait while it overthinks every decision.
The main problem I have with modern models is not that they fail to think. It is that they sometimes cannot stop.
Sol Against Fable
Fable felt very different when I last had a Claude subscription, and it still helps with bringing in customer websites and similar work. I keep coming back to OpenAI models because I love the Codex harness and genuinely enjoy talking to them.
They rarely stop at "Congratulations, you smashed it." They are more likely to add, "You smashed it, but do not assume that means the result will continue forever." Not everyone likes that. I do. It keeps me grounded.
Computer Use Is Outstanding
Sol's computer use is absolutely fantastic. The computer-use team behind Codex has built something exceptional.
I can ask it to log into a cloud dashboard, add environment toggles, deploy the changes, and monitor whether production actually comes up correctly. It does not merely click around. It follows the task through and reports what worked and what failed.
There have been mornings when I closed my laptop, left it alone, came back later, and Sol had finished the task with a useful summary waiting for me.
That changes the relationship with an agent. It stops feeling like autocomplete and starts feeling like something that can genuinely own a workflow.
Front-End Work Has Improved Massively
GPT-5.5 had many strengths. Front-end design was not one of them.
The interfaces it produced without a strong visual reference were often dreadful. The reliable trick was to give it a screenshot of the intended design. Once it had that reference, it could usually follow the direction well enough.
GPT-5.6 is a massive step up. It has a better sense of hierarchy, spacing, and what a finished interface should feel like. It still needs direction, and I would never let any model invent an entire product identity without review, but the starting point is far stronger.
Swift Is Better, but Not Perfect
JavaScript is everywhere, so every major model has seen an enormous amount of it. Swift is different. The framework behaviour can be less obvious, and small native UI details can waste a surprising amount of time.
Sol is noticeably better with Swift than 5.5. It understands more of the native patterns and gets closer to the correct solution on the first attempt.
It is not flawless. I still had one of those sessions where I kept asking why a border was visible, only to research it myself and discover that the fix was simple. The model had been looking in the wrong place.
That is the reality of working with any agent. A stronger model reduces the number of those moments. It does not eliminate them.
The Verdict
Sol is a really strong model, and I love working with it. It asks questions that improve the product, follows plans through to completion, uses a computer exceptionally well, and is significantly better at user-interface work.
I also tested Terra. It was decent and came close to the GPT-5.5 level, although it did not quite reach it for me. Luna has impressed me as a very capable smaller model.
GPT-5.5 remains special because it changed how fast a reasoning model could feel. GPT-5.6 Sol takes that foundation and turns it into something more dependable.
I cannot wait to see what GPT-6 looks like.