OpenAI Codex Is Retiring the Last Model that Behaves Like an Assistant

I opened Codex CLI today and it showed me this notice: "GPT-5.5 retires on October 14, 2026. Switch to GPT-5.6 Sol to continue working in Codex." GPT-5.5 is my main driver. I picked it over every newer model OpenAI has shipped since, and now I have just over three weeks left with it.

OpenAI's model docs say GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on all plans on October 14. The API is not affected. That does not help me. I run Codex on two $200 subscriptions, and I am not going to pay API prices on top of $800 a month in subscriptions. GPT-5.5 launched on April 23. Launch to retirement is 174 days.

I wrote in August that switching from GPT-5.5 to GPT-5.6 made me less productive, and after a week with GPT-6 Astra I went back to GPT-5.5 again. The reason is that GPT-5.5 is the last model that behaves like an assistant. By default it works with you. You discuss a task, it does it and comes back, you look at the result and go on from there.

Model behavior is never black and white, and GPT-5.6 Sol can do the same thing, but usually it does not. Sol has its own agenda. I have had Sol sessions that ran for 24 hours or longer, or until the usage was gone, on tasks GPT-5.5 would have handed back after an hour. In my Codex logs, Sol used 2.25x the tokens per session of GPT-5.5. Astra is the same family one level up. Both are more intelligent than GPT-5.5, both are hard to work with, and both eat the subscription. Luna is okay for simple tasks and automations, but not good enough as the main driver for software engineering. Terra I would never use, and I have written why. OpenAI's own docs describe Terra as "competitive with GPT-5.5 at a lower cost", and still the migration notice sends GPT-5.5 users to Sol, the model I measured at more than twice the tokens.

My setup until today was GPT-5.5 as the main driver all day, with Sol, Astra or occasionally Fable for code reviews and second opinions.

I am not the only one. On September 15 a Pro subscriber asked OpenAI in the developer forum to keep GPT-5.5 as a legacy option in Codex, even with usage caps. The replies say what I have been saying since August. Sol over-reasons, expands the task and burns quota. The OpenAI staff replies in the thread point to the API and suggest benchmarking GPT-5.6 Luna.

So the obvious move is Claude Code, and I will probably use it a lot more. I am not happy about that either. Fable 5.1 is an excellent model and I would use it all day for everything, but Anthropic only lets you put half of the Max plan's weekly usage into Fable, and Fable uses that up faster than the other Claude models. For me that is about half a day of work. Opus 5 gets work done, but I do not like the way it works, how it talks and I'm never sure if it's just hallucinating.

From the outside it looks like OpenAI has bigger problems than my workflow. On September 10 it paused new sign-ups to the $200 Pro plan because it cannot serve the demand for Astra. Now it retires the one subscription model that used my quota efficiently. I will not be completely off Codex. But from October 14 most of my daily work probably will move to Claude Code.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com