Codex: pricing, limits, errors and comparisons
Codex is OpenAI's coding agent, used from the CLI, IDE extensions, the desktop app and ChatGPT.
27 articles · last updated Nov 14, 2026
Free tiers and limits
- ChatGPT and Codex Shared Usage Limits in December: What Counts
Codex and Work already share one allowance. December adds non-Instant ChatGPT chat on Plus, Pro and Business; the start day and pool size are still unpublished.
Updated
- Codex 429 After Retries: Find the Limit That Owns the Request
The message says Codex exhausted its automatic retries after repeated 429 responses. It does not identify the quota owner, so recover by finding the active route before retrying again.
Updated
- Are Codex Limits Shared Across Accounts? Check the Account, Workspace, API, and Device Meters
Separate normal Codex usage sharing from same-machine account-switching symptoms with a meter board, evidence checklist, and safe continuation choices.
Updated
Errors and fixes
- Failed to Start Codex App-Server or “Not Available”: Fixes by Error Text
The exact text decides the fix: repair a path an update broke, lighten a slow start, find the crash, update the CLI to 0.162.1 or see if only cloud tasks fail.
Updated
- Codex "The browser access host is unavailable": No CLI Fix Yet
Terminal and IDE sessions have no fix yet, so do the browser step in a desktop app chat. If the app itself or @Chrome fails, follow OpenAI's checks in order.
Updated
- Codex Windows Sandbox Setup Fails with os error 32: Fixed Builds
Windows won't let the sandbox helper open a file Codex itself is running. Put the app and every CLI on a fixed build, clear leftovers, then check for errors=[].
Updated
- Codex Update Check Failed (HTTP 520): Fixed Server-Side on Oct 21
Not your Mac: OpenAI's update feed returned HTTP 520 from October 13 to 21, 2026. Quit and retry; if it still fails, reinstall or look for a Managed policy.
Updated
- OpenAI Developers "OAuth state not found": Connect From the Web
The desktop app's plugin callback fails on OpenAI's side, so retrying won't help. Pick a route by what you were connecting, with the evidence behind each one.
Updated
- Codex Cannot Serialize Value: Worktree Chats Fail, Local Works
Restarts won't clear it: build 26.1016.51333 breaks only Worktree chat starts. Local sends fine, and Hand off or your own git worktree restores isolation.
Updated
- Codex Timed Out Loading the Cloud Config Bundle or Failed to Load It
A timeout most often traces to a stalled network route; failed-to-load to repeated errors such as a revoked sign-in or an untrusted certificate. Logs decide. Keep .codex intact.
Updated
- Codex Keeps Reconnecting? Find the Cause Without Losing Your Task
Five reconnect attempts do not identify the cause. Save the prompt and changed-file state first, then use one new task, one basic command, and one controlled network comparison to narrow the failure.
Updated
- Codex Remote Compact Failed: Read the Suffix Before You Retry
The wrapper tells you that conversation compaction failed. The text after it tells you whether to inspect the stream, route, parameter, context, policy, parser, or child process.
Updated
Comparisons and alternatives
- Codex Plus vs Pro 5x vs Pro 20x: Pay for the Limit You Actually Need
Plus is the sensible default, Pro 5x buys sustained headroom, and Pro 20x is for people who can prove they will use it. The old 10x label was a temporary promotion, not today's third plan.
Updated
- Claude Code vs Codex in 2026: Which Coding Agent Should You Use Now?
Claude Code still fits supervised local coding best; Codex fits delegated multi-surface engineering work. The practical choice depends on where the job should run and who owns review.
Updated
- GPT-5.4 mini vs GPT-5.3-Codex: Which Model Should You Use?
GPT-5.4 mini is the better default for new API and subagent work, but GPT-5.3-Codex still has a real place in terminal-heavy coding and Codex cloud-task workflows. This guide explains exactly when to choose each.
Updated
- Gemini 3.1 Pro vs Opus 4.6 vs Codex 5.3: The Definitive Comparison (March 2026)
Three frontier models released in February 2026 have fundamentally changed the AI landscape. This guide compares Gemini 3.1 Pro, Claude Opus 4.6, and GPT-5.3-Codex with browser-verified pricing, real benchmark data, and practical recommendations for developers choosing their next production model.
Updated
Guides and setup
- Codex Token Exchange Failed: Fix "error sending request for url" at Sign-In
Sign-in could not be completed or Login server error? Codex itself cannot reach OpenAI. Check its proxy, the Windows app's proxy, firewall, then network.
Updated
- ChatGPT Go Codex Usage Explained: Monthly Limits, What You Get, Upgrade Paths
Go includes Codex on a monthly allowance OpenAI doesn't size. Check Settings → Usage or /status, then wait, use a reset, add credits if shown, or move to Plus.
Updated
- Codex “Responses WebSocket messages must have type”: Router Fix
Older codex-router builds reject the response.interrupt Codex sends mid-turn. Update the checkout to main; as of October 22, 2026, Homebrew still ships v0.6.0.
Updated
- Codex High CPU Usage: Match the Hot Process to the Right Fix
Busy while it works is normal. Busy while idle, after an update or in a system process is a bug: find the hot process and use the fix OpenAI or users confirmed.
Updated
- Codex self signed certificate in certificate chain: 4 Fixes
The certificate is almost always your company's root CA. Codex's client, the desktop app, MCP servers and sandboxed commands each read a different setting.
Updated
- How to Generate and Edit Images in Codex with GPT Image 2
Ask Codex for the image you need, refine it with a reference, and bring the saved result into your project. The key is to request an actual generated file and keep image usage separate from API billing.
Updated
- Connect Codex to a Third-Party Model API Without Guessing the Protocol
An OpenAI-compatible label is not enough: Codex needs a Responses-compatible route, a real provider model ID, working streaming, and credentials owned by the same route.
Updated
- Does Codex Require a Phone Number to Sign In?
A phone number is not universal for Codex, but an official ChatGPT sign-in can still redirect some accounts to add-phone. The screen you see determines whether you can continue, change auth routes, or need OpenAI Support.
Updated
- Codex Subagents: A Practical Guide to Delegation That Holds Up
Subagents work best when the branches are genuinely independent, the return contract is explicit, and one main thread still owns the final decision.
Updated
- Codex Sandbox and config.toml: Permissions Without Full Access
The sandbox controls where commands can act; approvals control when Codex must stop and ask. Diagnose that boundary before granting broader host access.
Updated
- Azure OpenAI for Codex: Setup Route, Surface Limits, and Troubleshooting
Codex can run through an Azure OpenAI provider route for CLI, IDE, and automation work, but deployment names, env vars, Entra ID, direct OpenAI API-key usage, and Codex cloud features are separate decisions.
Updated