
GPT-5.5 Retirement Hits Codex on October 14
OpenAI is retiring GPT-5.5 from ChatGPT, ChatGPT Work, and Codex on October 14, 2026, so hardcoded model strings in CI, cron jobs, and Codex configs need a quick update to gpt-5.6-sol. The episode also covers the removal of the legacy MCP server command and a handy new Codex desktop hotkey for inserting reusable prompt text.
Chapter 1
The Model Retirement Deadline and Migration to GPT Five Point Six Sol
Maya
So, October fourteen, twenty twenty six. That date is officially etched on my calendar now, and it really should be on yours too if you use Codex.
Ethan Park
Yeah, OpenAI put out a pretty direct announcement. GPT five point five is getting retired across ChatGPT, ChatGPT Work, and Codex.
Maya
Every plan, right? Consumer, Business, Enterprise, Edu, all of them.
Ethan Park
Every single one. Quick shoutout to Jellypod, by the way, for keeping this daily update running so we can digest these changes as they drop! But yeah, if you have automated scripts running in CI or scheduled cron jobs with hardcoded model strings, come October fourteen, those calls are just going to throw fatal model errors.
Maya
Fatal errors in headless scripts are... well, they're the stuff of pure Monday morning nightmares.
Ethan Park
So what is the exact swap? If you're signed in to Codex with a ChatGPT account, you need to change any explicit flag or default setting pointing to gpt five point five over to gpt five point six sol.
Maya
GPT five point six Sol. Got it. Wait, um, does this affect the standard OpenAI API key users too?
Ethan Park
No, OpenAI explicitly noted that this retirement doesn't apply to the direct OpenAI API, just ChatGPT sign-in sessions in Codex and ChatGPT products.
Maya
Okay, so how do people actually go hunt these down before things break? Because I know from experience, hardcoded strings love to hide.
Ethan Park
Start simple in your terminal. You can run grep, dash... I mean, grep r n with gpt five point five inside your dot codex directory. Check your repo level codex dot json files, your custom agent YAML definitions, and any CI environment variables like CODEX MODEL.
Maya
You almost said the word dash! Good save. But seriously, I was doing an audit on some older build pipelines last week, and I found a three month old cron job that was still silently calling an older model fallback.
Ethan Park
Right, because while it's running, it works fine and nobody notices. But the minute the deprecation date hits...
Maya
Poof. Complete silent failure or broken builds. It's those background, headless tasks that get you every single time.
Chapter 2
Protocol Cleanup and the Prompt Key Upgrade
Ethan Park
And speaking of cleanups and things going away, we need to talk about MCP.
Maya
Ah, right! The codex mcp server command?
Ethan Park
Gone. Following the deprecation notice back in late August, both the codex mcp server command and the standalone codex mcp server binary have been completely removed in the September releases.
Maya
So if someone was running a legacy integration that spun up that standalone binary, what's the path forward?
Ethan Park
You either connect to external MCP servers using codex mcp, or move to the experimental codex app server command if you were using it as a host service.
Maya
Okay, clean break. Now, on a lighter note, there was a really neat little quality of life feature added to the desktop app recently, right? Something about a hotkey?
Ethan Park
Yes! The Codex Micro key assignment. You can now map a hotkey specifically to Insert text.
Maya
Wait, how is that different from just pasting something into the prompt composer?
Ethan Park
It lets you inject reusable prompt snippets or template text right into your active turn, but without immediately firing off the send command! So you can insert a complex system instruction, edit a parameter, and then hit Enter when you're ready.
Maya
Oh, I love that. When you're testing agent behavior, accidentally auto sending an incomplete prompt is so frustrating.
Ethan Park
Totally. But listen, bringing this all back to the model migration, there is one big operational trap people need to test right now.
Maya
What's the trap?
Ethan Park
When you switch long running agent threads from gpt five point five to gpt five point six sol, tool call formatting and token context boundaries can behave slightly differently. If you just change the string on October thirteen and hope for the best...
Maya
You might find out your complex multi step tool loops break in production on October fourteen.
Ethan Park
Exactly. Spin up non production worktrees today, point them to gpt five point six sol, and run your full test suites against the new model now.
Maya
Test it early, fix your config files, and clean out those old MCP binaries. Good advice. Good chatting, talk soon!