Codex Automations Silently Default to GPT-5.2 — Here's How to Fix the Hidden Model Setting
Codex automations quietly use GPT-5.2 instead of GPT-5.5 by default. This hidden setting caused a 40-minute automation to stall. Here's the fix.
Codex Automations Were Running on the Wrong Model the Whole Time
Codex automations silently default to GPT-5.2, not GPT-5.5. If you’ve set up a scheduled automation in the Automations tab and never manually changed the model setting, it has been running on a model two generations behind what you’re using in your main chat sessions. This is the kind of default that doesn’t announce itself — you just notice your automation is slower than expected, or stalls entirely, and you spend 40 minutes debugging the wrong thing.
That’s exactly what happened in a documented build session: a weekly YouTube comment analysis automation — pulling 200 comments, running analysis, updating an Excel workbook, and pushing changes to a GitHub repo that auto-deploys via Vercel — ran for 40 minutes and stalled. The main chat session using GPT-5.5 on high reasoning had completed the same workflow in roughly 20 minutes during the initial build. The automation was doing the same work, slower, on a worse model, and nothing in the UI flagged it.
The fix is straightforward once you know where to look. But the fact that you have to know to look is the problem.
Why This Isn’t Obvious
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
The Automations tab in Codex is separate from your chat interface. When you set up a scheduled automation — essentially a cron-style task that injects a prompt into a new Codex chat on a schedule — the model configuration for that automation is independent of whatever model you’ve selected in your active session.
Your main chat might be running GPT-5.5 on high intelligence with standard speed. Your automation has its own model, reasoning level, and speed settings. And those settings default to GPT-5.2.
There’s no warning. No banner that says “this automation will run on a different model than your current session.” The automation just runs, and if you’re not watching closely, you might not notice the model version in the automation’s settings panel.
The broader issue is that Codex has a lot of per-context model configuration. You set the model in the main chat toggle (GPT-5.5, GPT-5.4, etc.), the intelligence level (low / medium / high / extra high), and the speed (standard / fast). Each automation has its own copy of all three settings. The defaults for automations weren’t set to match whatever you’re using interactively — they were set to GPT-5.2, which at this point is a noticeably less capable model for agentic tasks.
This is a real performance gap. GPT-5.5 is meaningfully more token-efficient than older versions, which matters a lot for multi-step agentic workflows. In the same build session, the creator noted that Codex sessions were lasting significantly longer than equivalent Claude Code sessions, attributing it specifically to GPT-5.5’s efficiency on both input and output tokens. Running an automation on GPT-5.2 instead throws away that efficiency advantage. If you’re curious how GPT-5.5 stacks up against Claude on coding tasks specifically, the GPT-5.5 vs Claude Opus 4.7 coding comparison covers the token efficiency difference in detail.
Reconstructing What Actually Happened
The automation in question was built to run every Sunday at 5:00 p.m. It was supposed to: invoke the YouTube comment insights skill, pull fresh comments via the YouTube Data API v3, update an Excel workbook with new rows, verify the dashboard, and commit changes to the GitHub repo so Vercel would auto-deploy the updated dashboard.
The first run stalled for a different reason — the Excel file was open locally, which blocked the write operation. That’s a user error, easy to fix. But after closing the file and re-triggering the automation, it ran for another extended stretch before the operator stopped it and asked why it was taking so long.
That’s when the GPT-5.2 default was discovered.
The automation settings panel shows the model, reasoning level, and speed. The model was set to GPT-5.2. After manually changing it to GPT-5.5 with high reasoning, the same automation completed in approximately 7 minutes.
That’s a significant difference for what should be a routine data refresh task. The workflow wasn’t complex — it was running a skill that had already been built and tested, hitting an API endpoint it had already authenticated against, writing to a file format it had already generated. GPT-5.5 handled this efficiently. GPT-5.2 apparently got stuck on something in the process.
The exact failure mode wasn’t fully diagnosed, but the pattern is consistent with what you’d expect: a less capable model in an agentic loop has more trouble recovering from minor obstacles, takes more steps to accomplish the same task, and burns more session time in the process. For a workflow that runs unattended, that’s a compounding problem — you’re not there to steer it back on track.
How to Actually Fix It
The fix requires going into each automation individually and updating three settings:
Model: Change from GPT-5.2 to GPT-5.5 (or GPT-5.4 if you have a reason to prefer it).
Reasoning: Set to high for most agentic workflows. medium is fine for simpler tasks like status reports or file reorganization. extra high is rarely necessary and costs more session usage.
Speed: Leave on standard unless you have a genuine time constraint. fast burns session usage faster and the speed difference rarely matters for scheduled automations running in the background.
You can view and edit these settings by clicking into the automation in the Automations tab. The prompt that gets injected into the new Codex chat is also visible there — worth reviewing to make sure it’s still accurate if you’ve iterated on the underlying skill since setting up the automation.
One thing to check while you’re in there: the automation prompt should reference your skill by its slash command name (e.g., the YouTube comment insights skill would be called via /youtube-comment-insights or however you named it when Codex created it). Skills live globally in ~/.codex/skills/ or locally in your project directory, and the automation needs to be able to find them. If you’ve moved or renamed a skill since creating the automation, the prompt may be referencing something that no longer exists at that path.
The Deeper Issue: Per-Context Model Configuration
This GPT-5.2 default is a specific instance of a broader pattern in Codex: model configuration is per-context, and the defaults don’t always match what you’d expect.
The main chat toggle, the automation settings, and the /model slash command all control model selection in different scopes. If you use /model in a chat to switch to GPT-5.5, that doesn’t persist to your automations. If you set up an automation while your main chat was on GPT-5.4, the automation might have inherited that setting — or it might have defaulted to GPT-5.2 regardless.
The /reasoning slash command similarly controls reasoning level within a chat session, but automations have their own reasoning setting that’s independent.
This is worth understanding because it affects more than just automations. The agents.md file (Codex’s equivalent of Claude Code’s CLAUDE.md) is read at the start of every new chat, including the chats that automations spawn. If you want your automations to use specific model settings consistently, you can document that preference in agents.md — something like “always use GPT-5.5 with high reasoning for any automated workflow” — and Codex will factor that into its behavior when the automation chat initializes.
It’s not a substitute for checking the automation settings directly, but it adds a layer of consistency. The Claude Code Skills vs Plugins post covers a similar principle for Claude Code: the difference between what’s baked into a skill versus what’s configured at runtime. The same distinction applies here.
What This Means for Automation Design
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
If you’re building automations in Codex — scheduled data refreshes, weekly reports, anything that runs unattended — treat model selection as a first-class configuration decision, not an afterthought.
The practical checklist when setting up any new automation:
-
Set the model explicitly — don’t trust the default. GPT-5.5 for anything involving multi-step agentic work. GPT-5.4 if you’re on a tighter session budget and the task is simpler.
-
Set reasoning to
highfor workflows that involve file operations, API calls, or any conditional logic.mediumis fine for pure text generation tasks. -
Test the automation with
run nowbefore relying on the scheduled trigger. Watch the spawned chat. If it’s taking longer than your manual build session took, something is wrong — check the model setting first. -
Keep the Excel file (or any output file) closed when the automation is scheduled to run. This sounds obvious but it’s the kind of thing that bites you at 5:00 p.m. on a Sunday.
-
Check the automation prompt after any skill updates. Skills are just markdown files, and if you’ve iterated on the skill recipe since creating the automation, the automation’s injected prompt might be out of sync.
The rate limits panel is also worth monitoring. Codex shows a 5-hour session reset and a weekly reset, with percentage remaining visible. Automations consume session usage just like interactive chats do. Running a poorly-configured automation on GPT-5.2 that takes 40 minutes instead of 7 minutes isn’t just slower — it’s burning more of your weekly session budget on a task that should be cheap.
For teams building more complex orchestration — chaining multiple automations, routing between models based on task type, connecting to business tools — platforms like MindStudio handle this at a different layer: 200+ models, 1,000+ integrations, and a visual builder for composing agents and workflows without managing per-context model configuration manually.
Skills, Automations, and the Portability Question
One thing worth understanding about Codex skills: they’re just markdown files. The global skills directory at ~/.codex/skills/ is readable by any agent harness that knows to look there — Claude Code, Cursor, OpenClaw, Codex itself. If you build a skill in Codex and later want to run it from Claude Code, you don’t have to rebuild it. You might need to rename the file or adjust the frontmatter slightly, but the recipe is portable.
This matters for automations because it means the skill underlying your automation isn’t locked to Codex. If you decide to move a workflow to a different harness, or if you want to run the same skill from multiple tools, the markdown file is the source of truth. The automation in Codex is just a scheduled trigger that injects a prompt referencing that skill.
The /skill-creator slash command in Codex can reverse-engineer a skill from a completed workflow — you run through a task manually, get an output you like, then ask Codex to turn that into a skill. That skill then lives in the global directory and is available to any automation you set up. It’s a clean workflow: build interactively, extract the recipe, automate the recipe, check the model settings before you walk away.
One coffee. One working app.
You bring the idea. Remy manages the project.
On the question of going from a skill-backed automation to a fully deployed application — the YouTube analytics demo in the source material went from Codex automation to GitHub repo to Vercel deployment — the abstraction level keeps rising. Tools like Remy take a different approach to that final step: you write a spec in annotated markdown, and it compiles into a complete TypeScript stack with backend, database, auth, and deployment. The spec is the source of truth; the generated code is derived output. Different model, different use case, but the same underlying question: where does the authoritative description of what you’re building actually live?
The Takeaway
The GPT-5.2 default in Codex automations is a bug, or at minimum a misconfigured default that OpenAI will presumably fix. But until they do, every automation you’ve set up without manually changing the model is running on GPT-5.2.
Go check your automations now. Click into each one, look at the model setting, change it to GPT-5.5, set reasoning to high, and re-run with run now to verify the timing is what you’d expect. It takes two minutes and it will materially change how reliably your automations complete.
The broader lesson is that in any agentic system with per-context configuration, the defaults are load-bearing. They’re not neutral. A default of GPT-5.2 in automations means every user who doesn’t know to check is running degraded workflows. Checking model configuration should be part of the setup checklist for any automation, the same way you’d check that your API keys are in .env.local and not committed to the repo.
For more on how model choice affects agentic task performance — particularly the cost and speed tradeoffs between GPT-5.4 Mini and Claude Haiku for sub-agent work — the GPT-5.4 Mini vs Claude Haiku sub-agent comparison is worth reading alongside this. And if you’re running similar scheduled workflows in Claude Code, the token-saving approach using Opus plan mode covers the equivalent configuration decisions on that side of the fence.
