OpenAI Codex gpt-5.4 'Not Supported' Error

The Exact Error and Who Got Hit

Starting in March 2026, users across multiple plans began reporting this exact error message in both the Codex web app and the CLI:

Error: The 'gpt-5.4' model is not supported 
when using Codex with a ChatGPT account.

What made this particularly frustrating: the model still appeared in the dropdown menu. You could select it. You just couldn't use it. The UI and backend were out of sync.

Affected users included:

• ChatGPT Free and Go plan users (expected — gpt-5.4 full isn't available on free tiers)
• ChatGPT Plus subscribers ($20/month) — not expected, this is a bug
• Business and Enterprise users — also affected intermittently
• Codex CLI users authenticated via ChatGPT login (not API key)

Why This Error Happens: 4 Root Causes

Based on confirmed reports from the OpenAI Developer Community and GitHub issues, there are four distinct causes:

1. Backend Model Routing Regression

OpenAI's internal model router sometimes misidentifies which models your account can access. This is the most common cause and typically resolves itself within 24-48 hours as OpenAI pushes fixes. But if you can't wait, the manual fixes below will get you working immediately.

2. ChatGPT Login vs API Key Authentication

This is the root cause most people miss. Codex distinguishes between two authentication methods:

• ChatGPT account login: Uses your consumer subscription (Free/Plus/Pro) — some models are restricted
• API key authentication: Uses your developer account with credit balance — broader model access

Some models, including certain gpt-5.4 variants, are gated behind API-only access and will throw errors when you try to use them through a ChatGPT login, even if you're a paying subscriber.

3. Model Name Mapping Bug in CLI

A specific bug in certain Codex CLI versions causes the internal model name to get rewritten. For example, gpt-5.4 gets incorrectly mapped to gpt-5.4-t, which doesn't exist — resulting in a 404 error during startup. This is a known issue documented on GitHub.

4. Stale Session / Cache Issue

Your local Codex session may have cached old model permissions. If OpenAI updated model access rules server-side, your local cache could still reflect the old permissions — showing the model as available but blocking its use.

Step-by-Step Fix Guide

Try these in order. Most users get unstuck by Step 2 or 3:

Fix 1: Update Codex CLI to Latest Version

Always start here. The model mapping bug was fixed in recent releases:

# Update Codex CLI to latest
npm install -g @openai/codex@latest

# Verify version (should be 0.122.0+)
codex --version

Fix 2: Switch Authentication to API Key

If you're logged in via ChatGPT account, switch to API key authentication:

# Log out of ChatGPT account
codex logout

# Set API key instead
export OPENAI_API_KEY="sk-your-api-key-here"

# Restart Codex
codex

Important: API key access requires credit balance in your OpenAI developer account. But it gives you broader model access than ChatGPT login.

Fix 3: Clear Local Model Cache

If Codex is using stale model permissions:

# Find and delete the models cache
rm ~/.codex/models_cache.json

# For VS Code extension: 
# Settings → Extensions → Codex → Sign Out
# Restart VS Code → Sign back in

Fix 4: Check Your config.toml

Open your Codex config file and make sure the model name is correctly specified:

# ~/.codex/config.toml

# ❌ Wrong (may cause mapping bug)
model = "gpt-5.4"

# ✅ Correct (explicit full name)
model = "gpt-5.4-thinking"
# Or for Codex-optimized:
model = "gpt-5.3-codex"

Fix 5: Verify Your Plan Includes gpt-5.4

Not all plans include the full gpt-5.4 model in Codex:

3 Fallback Workarounds If Nothing Works

If the fixes above don't solve it (sometimes OpenAI's backend is simply broken), use these workarounds to keep coding:

Workaround 1: Switch to gpt-5.4-mini

The mini variant is available on all plans (including free) and is surprisingly capable. CNET reports it's "more than twice as fast as its predecessor" with near-gpt-5.4 level capability on coding tasks. Not ideal for complex multi-file refactoring, but solid for day-to-day coding.

Workaround 2: Use gpt-5.3-codex as Fallback

The gpt-5.3-codex model is still available and was specifically optimized for code generation. Update your config:

# In config.toml
model = "gpt-5.3-codex"

Workaround 3: Switch to Claude Code Temporarily

If you need to keep working and Codex is completely broken, Claude Code is a strong alternative. Anthropic's Claude Opus 4.7 currently leads on SWE-bench for agentic coding — we covered this in our Cursor vs GitHub Copilot vs Claude Code deep dive. OpenAI even built a codex-plugin-cc bridge that lets Codex commands run inside Claude Code — that's how seriously developers take the interoperability between these tools. For context, Claude's own literal mode changes caused similar developer frustration.

Prevention: Stop This From Happening Again

• Pin your model version — Always specify the exact model name in config.toml instead of relying on "latest"
• Use API key auth — More reliable than ChatGPT login for developer workflows
• Keep CLI updated — Run npm update -g @openai/codex weekly
• Monitor the status page — Bookmark status.openai.com for outage alerts
• Watch GitHub Issues — The openai/codex repo is where bugs get reported first
• Have a fallback ready — Always keep gpt-5.3-codex or Claude Code configured as backup

Key Takeaways

• The "gpt-5.4 not supported" error is a known issue affecting multiple plan tiers
• Four root causes: backend regression, auth method mismatch, CLI mapping bug, stale cache
• Try fixes in order: update CLI → switch to API key → clear cache → check config → verify plan
• gpt-5.4-mini and gpt-5.3-codex are reliable fallbacks when gpt-5.4 is unavailable (see our 10 best coding AI 2026 roundup)
• Claude Code is a viable alternative with top-tier coding benchmarks — see our agentic coding tools guide
• Pin model versions and use API key auth to prevent future occurrences

Frequently Asked Questions

Why does Codex say gpt-5.4 is not supported?

This error occurs when Codex can't route your request to the gpt-5.4 model. Common causes include using a ChatGPT login instead of an API key, being on a plan that doesn't include gpt-5.4 access, a model name mapping bug in older CLI versions, or a backend routing regression on OpenAI's side.

How do I fix the Codex model disappeared bug?

Update Codex CLI to the latest version, switch from ChatGPT login to API key authentication, clear your local models_cache.json file, and verify your config.toml has the correct model name. If gpt-5.4 still doesn't work, use gpt-5.4-mini or gpt-5.3-codex as fallbacks.

Which ChatGPT plans include gpt-5.4 in Codex?

ChatGPT Plus ($20/month) includes gpt-5.4 Thinking in Codex with usage limits. The Pro plans ($100 or $200/month) include gpt-5.4 Pro with higher Codex usage allowances. Free plan users only get access to gpt-5.4-mini. API key users can access all models with pay-per-token pricing.

Is gpt-5.3-codex a good fallback for gpt-5.4?

Yes. The gpt-5.3-codex model was specifically optimized for code generation and remains available as a stable fallback. It won't match gpt-5.4's reasoning depth, but for day-to-day coding tasks like debugging, refactoring, and code review, it performs reliably.

Can I use Claude Code instead of OpenAI Codex?

Yes. Claude Code powered by Anthropic's Claude Opus 4.7 currently leads on SWE-bench for agentic coding benchmarks. We recently tested Claude 4.7's 3.75MP vision upgrade too. OpenAI has even built an official codex-plugin-cc bridge for interoperability. It's a strong alternative when Codex is experiencing issues.

When was gpt-5.4 released for Codex?

OpenAI released gpt-5.4 on March 5, 2026, making it available in ChatGPT, Codex, and the API. The model is available as gpt-5.4 Thinking for Plus users and as gpt-5.4 Pro for Enterprise and Edu subscribers. It's priced at $2.50 per million input tokens in the API.