OpenAI Codex on X

The largest group is people showing what Codex built. They report vibe-coding a notes site, a first app from someone with no programming background, a 3D workflow that models parts from drawings and lays them on a print bed, and a two-to-three week job finished in a few days; one ran four large tasks in a day. Output gets called clean, and one user says Codex proposed a skill-tree editor they had not thought of.

Reservations sit on speed and clutter rather than the work: one calls voice input "a life changer" but wants latency much lower; another says the updates feel minor and throughput still trails Claude.

Disappointment runs along that same comparison: people say it is slower and burns more quota than Opus 5.5, cancel or let a subscription sit unused, and post that "Codex is dead". A near-equal group of 34 says the opposite — that it matches or beats Claude Code on their own work, better at compaction and at following instructions. 27 are happy with the quota and welcome the resets; 7 are still deciding.

The biggest complaint is what Codex does to the code: rewriting a file someone only wanted commented, adding screenshot modes nobody asked for ("sideways, rotating, vintage mode"), over-testing, leaving a bug half-fixed and blocking other templates, and using vague, non-committal language.

Quota is the second: thousands of credits burned on about fifteen throwaway tasks, half a token spend going on corrections, and small tasks eating a fifth to a third of a five-hour window where Claude Code used a tenth. Bugs and breakage come next — an update that rolled recent project work back in time, projects vanishing from the UI, one session file per session piling up, a login locked behind an SMS code that never arrives, and memory leaks.

Then the timing complaints: a hard five-hour stop mid-task, which Pro users voted to remove, and resets that land in the middle of the night, arrive unannounced, or leave saved resets expiring unused. 8 say the resets stand in for real improvements. Slower answers ("three hours on a job that takes one") and remote sessions that keep reconnecting — "its main feature now" — are level. Small groups flag bans for security work and worry about what Codex can reach.

Claude Code is the benchmark. Some use it instead: people downgrade a Codex plan after moving back to Claude, say they are leaving for it over token burn, or note that Opus 5.5 is why their Codex tab stays closed. Others run the two together, with Codex as orchestrator and Claude as the engineer, Claude building while Codex reviews ("codex keeps catching stuff claude missed"), or one thread routing work to whichever has headroom.

The cost argument cuts both ways. Another benchmark post has Claude Code with Opus 5.5 at 61.7 points and about $11.42 a task against Codex with Sol at 54.1 points and $1.50; others describe routing everyday work to Codex and hard problems to Opus. DeepSeek is named as the cheapest option, with Codex as one harness among several for it.

Cursor draws the "you already have Codex and Claude" reply; Grok and Muse are named by a few as writing better for them. A handful say Antigravity finishes work Codex stalls on, and Hermes and Pi are named as harnesses for the Codex models.

The how-to group is the largest by far. People ask how to install it and run /init to create AGENTS.md, how to hide models from the model picker, how to prompt a skill, how to get a PR-review prompt to work, and why a Datadog MCP shows "No data". Others answer with setup recipes and workflow tips: run it only when there is coding to do, branch to Work first, keep notes in AGENTS.md and symlink CLAUDE.md to it so both agents start from the same page. Several say they keep asking ChatGPT the same question: what is the difference between Work and Codex, and when should I use which?

Codex or Claude is the next question, often a budget call — "$200 for AI coding tools: Claude, Codex, Gemini, Cursor, where are you spending it?" — or a request for one solid reason to switch, or which of the two to trust to ship a production app.

On cost, people ask which tier a feature needs, whether Codex Plus and Claude Pro are the same price, and why they would pay for the $20 plan at all. A smaller group asks whether Codex can do a particular job — edit video, run from a phone — and whether a reset just landed.

The handoff request is concrete. People already run Claude and Codex side by side with automated handoffs, and one describes being "the mailman between claude and codex" because there is no signal that the other has finished. Others point at the thin part: what passes between agents is a summary, so whatever the first one missed the second never sees ("A 漏了什么，B 根本不知道"). They want Codex and Claude Code to work as equal peers over a long stretch, and to keep going with the same context when one hits its quota wall, instead of restating the whole task.

The rest are small, specific asks. One wants to choose between a ChatGPT or Codex chat from the desktop app, see recent chats in place, and manage several accounts; another wants the ChatGPT and Codex surfaces merged so capabilities and knowledge are not split in two. Two want a read-and-agree step before diffs apply. Single people want dictation in the terminal, banked resets used automatically before they expire, work resumed after a reset without asking, headless use, and the image cap lifted.