Codex Update

9 stable releases and a 28-day run: Codex's October, all in one post. Everything that shipped or changed in October 2026, checked against GitHub, OpenAI's changelog, help pages and official posts. Bookmark it for reference. 1. Biggest launches - GPT-6.1 Terra came to Codex on Oct 26 for Plus, Pro, Business, Enterprise and Edu. Tibo says it gives 2x the usage of 6.1 Sol. - GPT-6.1 Sol Ultrafast arrived Oct 8. OpenAI says it's up to 8x faster than Sol Standard. Pro $500 and eligible Enterprise and Edu only. - Sol's default context in Codex went from 272k to 500k tokens on Oct 11, with no long-context fees on a ChatGPT plan. - Tibo started a 28-day run on Oct 4: one clear improvement or a full limits reset every day. - Also shipped: composer predictions (beta, Pro), a re-shipped Codex cloud with Tailscale support, a new Windows sandbox built on Microsoft's MXC, editing earlier messages in the desktop app, and about 60% less app data on disk plus a Storage tab. 2. Every stable CLI release - 0.160.0 (Oct 1): start sessions outside a project, fewer SQLite stalls, Windows sandbox fixes - 0.160.1 (Oct 5): keeps Windows env vars for remote MCP servers - 0.161.0 (Oct 7): GPT-6.1 Sol becomes the default model, /mcp login, mic and speaker picker - 0.162.0 (Oct 8): managed Git worktrees, pin tasks with p, /copy, clickable links - 0.162.1 (Oct 9): fixes a TUI crash and background-server startup failures - 0.163.0 (Oct 15): searchable /config, steer mid-response, leader key in /keymap, /daemon restart - 0.164.0 (Oct 23): ten /presets slots, inline /model and /permissions, image previews - 0.165.0 (Oct 27): built-in Vertex AI provider, Terra in the catalog, memory v2 on by default - 0.166.0 (Oct 29): codex doctor --network-only, transcript search, codex queue --resume 3. Pricing, plans and limits - Full limit resets landed again and again during the run, including Days 6, 8, 10, 13, 15, 21 and 23. The run closes with banked resets you redeem yourself. - Auto-review ("Approve for me") stopped counting against usage (Day 2). Day 12 stretched usage with longer caching, and Day 16 made the cache keep with no expiry for logged-in users. - GPT-5.5 left Codex for ChatGPT sign-in on Oct 14. API-key sessions kept it. - About 40,000 accounts had limits that didn't match their plan. OpenAI fixed it and added credits. Pro 500 upgraders hit by a separate bug got 25,000 credits. - Codex for OSS grants moved up to the Pro 200 tier. - Oct 29: limits doubled for Pro 100, 200, 500 and Business Premium, and the planned Pro 200 cut was cancelled. - Pro and Business now get monthly API credits. They don't add to your Codex allowance. - New device limits cap how many paid accounts one machine can hold ("Device limit reached"). - Coming in December: ChatGPT and Codex share one token-based allowance on Plus, Pro and Business. 4. What people hit, and the fixes - Sudden usage drain (mid-month): a confirmed bug, now fixed. If it happens again, run /feedback in the affected session and check /usage daily. - "Selected model is at capacity" or 429s with quota left (GitHub #57084): no official fix yet. Switch with /model, then codex resume to continue. - VS Code extension 26.1021 won't start its UI: users report the pre-release build, rolling back with Install Specific Version, or launching with code --disable-gpu. - Worktree chats fail with "Cannot serialize value" on Windows and Linux: Local mode still works. - Max or Ultra missing from the reasoning slider (#58201): re-check it in Settings > General > Available reasoning efforts. - Usage rings gone from the desktop sidebar (#57656): use Settings > Usage or /status. - Security: don't launch Codex inside ~/Downloads, where Forever Security found a planted AGENTS.md. Don't install from reseller one-liners either. Symantec caught one turning the sandbox off. 5. Three tips worth keeping - Get a baseline before December's shared pool: /usage daily and /usage weekly in the CLI. - Pipe anything into Codex: npm test 2>&1 | codex exec "summarize the failures and propose the smallest fix" - Mind the AGENTS.md cap. Codex merges your AGENTS.md files and stops at 32 KiB (project_doc_max_bytes), so split rules into nested folders. Ask for November: publish the per-plan token allowance before the shared pool starts, so users can do the maths. Sources: github.com/openai/codex releases and issues, developers.openai.com/codex/…, help.openai.com, and posts from @OpenAI, @OpenAIDevs and @thsottiaux

Codex this week: 9 changes, 5 user pain points, all here. Oct 25 to Nov 1, checked against OpenAI's posts, help pages and GitHub. Releases 1. GPT-6.1 Terra rolled out in Codex for Plus, Pro, Enterprise and Edu (Oct 26). Tibo says it gives 2x the usage of 6.1 Sol. 2. CLI v0.165.0 (Oct 27): built-in Vertex AI provider, change reasoning effort mid-chat, memory v2 on by default. 3. CLI v0.166.0 (Oct 29): Terra in the Vertex AI and Bedrock catalogs, codex doctor --network-only, transcript search, and a fix for early compaction after resume or fork. 4. Text your dot over iMessage or RCS (Pro, US preview). Multi-account switching landed on mobile, with desktop "coming soon". Limits and pricing 5. Limits doubled for Pro 100, 200, 500 and Business Premium, and the Oct 30 Pro 200 cut was cancelled. 6. Pro and Business now get monthly API credits ($100 to $500 on Pro). They don't raise your Codex allowance. 7. Confirmed: from December, ChatGPT and Codex share one token-based allowance on Plus, Pro and Business. 8. New device limits: one machine can only hold so many paid accounts, or you get "Device limit reached". 9. Banked reset 1 of 2 is live. Redeem it in Usage settings or with /usage in the CLI. Incidents and security - Oct 27: GPT-6.1 Sol was degraded for about 78 minutes, then resolved. - Don't run Codex inside ~/Downloads (Forever Security found a planted AGENTS.md), and don't install it from reseller one-liners (Symantec caught one turning the sandbox off). - Forever Security says a local HTML page opened in Codex's built-in browser can read files on disk. Still unpatched as of Oct 29, per their write-up. What users are saying - 429 and "at capacity" errors with most of the allowance left, mostly VS Code on Linux. The busiest new GitHub issue this week (#57084, 87 comments). No fix yet. Switch models with /model, and codex resume picks the session back up. - VS Code extension 26.1021.22027 "could not start its user interface" on Linux and Windows, across several issues. Reported workarounds: the pre-release build, an older version, or launching with code --disable-gpu. - Usage rings vanished from the desktop app sidebar after recent updates (#57656, Windows and Mac). Check Settings > Usage or /status meanwhile. - Max or Ultra missing from the reasoning slider after 26.1027 (#58201). Fix confirmed by several users: Settings > General > Available reasoning efforts, re-check the level. - The shared December pool is the loudest topic on X, and a GitHub request to keep Plus allowances separate is gaining votes (#58530). What people are asking OpenAI for: publish how chat and Codex tokens will be counted before December, bring the usage rings back, post a status note when 429s hit accounts with quota left, and give a fix date for the built-in browser issue. Community shoutout @talkaboutdesign built agent-stat, which shows your local Codex usage, per-request cost and live limits in one place: npx agent-stat@latest Tip Hold your banked reset until you need it. It only redeems once an allowance is under 80% remaining, and it moves your weekly reset date to the day you use it. Sources: openai/codex releases and issues on GitHub, help.openai.com, @thsottiaux and @OpenAI on X

6/7 Workaround until then. Check the version actually moved: b=$(codex --version); codex update; a=$(codex --version); echo "$b -> $a" If it didn't, pin the release: npm i -g @openai/[email protected] With an age gate on, npm then fails loudly (ETARGET) instead of quietly.

3/7 Fixed in 0.166.0: the updater now saves the script first and only runs it if curl succeeded. Same dead-proxy test: curl (7), a clear Error line, exit code 1, version unchanged. Correct. Fix by etraut-openai; root cause traced by jiang1997 & abruption. github.com/openai/codex/pull…

2/7 The old bug (standalone installs, 0.165.0 and earlier): the updater ran curl ... | sh. If curl fails, sh gets an empty script, exits 0, and Codex prints success. I cut the network with a dead proxy: curl error (7), "🎉 Update ran successfully!", exit 0, still 0.165.0.

1/7 Half fixed: Codex's "Update ran successfully" message. 0.166.0 stops the standalone updater from claiming success when the download fails. On npm installs it can still say 🎉 when nothing changed, or after a downgrade. Tested on a clean Linux box, no login needed 🧵

507 👍: LSP in the CLI is Codex's top feature request. Two of the top five are about undoing work: - Bring back /undo (490) - /rewind to roll back both the chat and the code edits (230) Method: open enhancement issues in openai/codex, ranked by 👍, snapshot Oct 29. If one of these is yours, add your 👍 on the issue.

5/7 It looks load-sensitive. On a local 8-core Linux box the new build opened fine for me (UI mounted in 9.8 s). Others report it reproduces in dev containers too (MichaelDevine, SyncM1972 on GitHub). Quick retries that help some: disable/re-enable the extension (wrickout).

3/7 Why it happens: the extension gives its webview 30 s to report "ready", then shows the error. I unpacked both builds. That 30 s limit is identical. What grew is the webview: 539 MB / 19,524 files before, 723 MB / 22,544 files now.

2/7 The fix most people confirm: roll back to the last good build. code --install-extension [email protected] --force On VS Code 1.143.1 it installed in ~13 s and was marked pinned. Via GitHub user StevieDicks, who said reinstalling from the menu didn't work.

Codex CLI v0.164.0 cheat sheet: resume and fork sessions, piping into codex exec, profiles, worktrees, /compact, /init, /presets, and remote control. Save this one.

5/8 Second trap: if Codex lives under ~/.local or nvm, ":minimal" = "read" may not cover its own binary: bwrap: execvp …/codex: No such file or directory A read rule for the install dir fixed it. Filed by GitHub user alma-iceee (+8): github.com/openai/codex/issu…

4/8 Heads-up: codex doctor shows ✓ sandbox for the same broken profile. Its own note says "resolve paths only; read/write access is not tested." Smoke-test a new profile before trusting it: codex sandbox -P <name> -- true It surfaces the panic in under a second.

3/8 Workaround that worked: grant write on a directory, not a file ("docs" = "write"). Writes in docs/ succeed, while main.py and .git/config stay read-only. Trade-off: the whole folder is writable. If you need a tight scope, move the file into its own folder.