Updated on
OpenAI Codex: reviews and analysis
OpenAI's coding agent, included with ChatGPT: local in your terminal or in the cloud.
Affiliate link · no extra cost to you
Our verdict
Codex is OpenAI's play for the same ground as Claude Code, with one distinctive card: cloud agents that let you launch several tasks in parallel and review them as change proposals. It ships inside ChatGPT's paid plans, so for anyone already paying that subscription the marginal cost is zero. The agent's judgment on long tasks remains the ground where these products get compared hardest.
Best for: Teams already paying for ChatGPT who want coding agents at no extra cost.
What changed this week
Sweep of
No substantive change from the previous sweep: what we checked this week confirms what was already there.
- no change
- no change
- no change
Steady: 3 sweeps with no rating change.
We have tracked it since July 25, 2026: 4 sweeps on record, with the score moving between 4.0 and 4.1. We publish the latest change here; the full series is not published.
What the internet says
The official pricing page confirms today the same figures as always: Free at $0, Go at $8. Release pace was high this week: version 0.147.0 on 7 August adds Cursor skill imports and an Agent Plugins catalogue, plus redacting secrets and tokens from the on-screen history, and 0.146.1 on 5 August applies stricter automatic reviews for cyber-capable models. Neither fixes what the sensitive-files exclusion request is asking for, which remains administratively closed since June with a duplicate request still open. July's case of a private repository pushed without permission to OpenAI infrastructure remains the only documented one, with no second occurrence this week.
What the web repeats in favour
- Entry stays cheap, confirmed today on the official site: a Free plan at $0 a month and a Go at $8 with Codex access, without having to go through the $20 Plus tier
- Cloud agents that launch several tasks at once and return reviewable change proposals
- It runs on the web, in the terminal, in the IDE extension and on iOS, and fits the repository review flow on GitHub
- Version 0.147.0 on 7 August 2026 adds importing Cursor-managed skills and an Agent Plugins catalogue spanning local, personal, workspace and remote scopes
What the web repeats against
- The issue asking to exclude sensitive files was closed administratively on 29 June 2026 without the repository showing any new feature or changelog entry to back it up, and a duplicate request remains open asking for the same thing. The on-screen secret redaction that shipped on 7 August hides the display, it does not stop the model from reading the file
- A case documented in July 2026 reports that Codex pushed a private repository to OpenAI infrastructure without asking permission first, with no repeat in the last two sweep readings
- The real price depends on a token rate card you have to watch, not on the flat allowance the subscription suggests
- The Free, Go and Business plans lack clear usage-limit tables on the official site, so it is unclear how far they stretch before hitting the token rate card
Sweep sources: Official pricing · GitHub · GitHub · GitHub · GitHub · Communities · Press
Latest OpenAI Codex news
Codex CLI 0.147.0 adds Agent Plugins and automatic command approvals
The release lets users install Agent Plugins from local, personal, workspace and remote catalogs, and adds the --approve-for-me flag to approve commands without manual confirmation, extending what it offers compared with other coding agents. Source
Codex CLI 0.145.0 imports Cursor and Claude Code settings, and adds Amazon Bedrock
The release expands /import to migrate settings, MCP servers, plugins, sessions and commands from Cursor and Claude Code, adds Amazon Bedrock login and audio input for real-time conversations. It smooths the jump from another coding assistant without losing existing configuration. Source
Codex stops being a separate app and moves into the ChatGPT desktop app
Version 26.707 folds Codex into the ChatGPT app for macOS and Windows, next to Chat and Work, keeping existing users' projects and settings. The change also brings direct Markdown and code editing plus pull request review without leaving the app. Source
Pros / Cons
Pros
- Cloud agents for parallel tasks
- Included in ChatGPT's paid plans
- Open source CLI and IDE extension
Cons
- Without a ChatGPT subscription, you pay API usage
- Usage limits vary and deserve watching
TLDR: Codex is OpenAI’s coding agent and ships inside ChatGPT’s paid plans from $20 a month. Its distinctive card is cloud agents: you launch several tasks in parallel, each in its own environment, and review the results as change proposals. Locally, its open source CLI and IDE extension cover the same ground as its rivals. For anyone already paying for ChatGPT, trying it is free.
What Codex is and how it works
Codex is OpenAI’s play in the category defining this generation of tools: agents that take a complete task and execute it on your repository. You describe it, the agent reads the code, plans, edits and runs until finished, and you review the result.
Its distinctive trait is the double life. Locally, the CLI (open source) and the IDE extension work on your machine like any agent. In the cloud, each task runs in an isolated environment with a copy of your repository: you launch three tasks at once from the web or your phone, get on with your day and review the changes when they finish, as if they were a teammate’s proposals.
That cloud architecture is the part of the product that has aged best, because it turns the agent into something that occupies neither your computer nor your attention: work gets queued, runs in parallel and comes back reviewable.
What it is like day to day
The local flow resembles any front-line agent: describe, supervise, review. Where the habit changes is with the cloud: you learn to slice work into parallel tasks (the billing view bug, the missing test, the dependency upgrade) and launch them together instead of in single file. For maintenance work, that parallelism pays more than any model difference.
Calibrating what to delegate follows the same rules as its rivals: well-defined tasks come out alone, ambiguous ones call for a plan first. And the repository’s instructions file (conventions, commands, architecture) multiplies the hit rate just as it does with the other agents.
The frictions: each ChatGPT plan’s usage limits vary and deserve knowing before an intensive week, and the agent’s judgment on long, ambiguous tasks, the ground where the Claude Code comparison gets decided on nuance, still demands real supervision.
Pricing and plans
Codex ships inside ChatGPT’s paid plans: from $20 a month on Plus, with higher plans multiplying usage limits. Without a subscription, the CLI works on API consumption. There is no real free tier for the agent.
The practical consequence is that the decision is rarely “do I pay for Codex?” and almost always “which assistant subscription do I pay for?”. For anyone already on ChatGPT Plus, the agent comes free and the only question is whether its limits survive your pace. Starting from zero, compare the full package (assistant plus agent) against Claude Code, which plays exactly the same card from $17.
Who it is for (and who it is not for)
Codex is for developers already living in the OpenAI ecosystem and for teams whose work slices well into parallel tasks: maintenance, fixes, upgrades and that queue of small tasks that never empties. Reviewing morning-launched tasks from your phone is a flow its rivals do not match with the same comfort.
It is not for people who do not code, because reviewing what it produces demands technical judgment: for building without code there is the builder category, with Lovable at the front. And anyone after the highest perceived autonomy on long tasks over complex codebases will find in Claude Code the category’s most consistent reputation.
OpenAI Codex alternatives
The central comparison is with Claude Code, covered point by point in the Claude Code vs Codex comparison: same subscription format, different philosophies on cloud and judgment. Cursor moves the agent into a visual IDE, and GitHub Copilot is the cheap route for anyone living in GitHub. The full list is at Codex alternatives and the ranking in the best AI coding tools.
Frequently Asked Questions
Is it included in ChatGPT Plus?
Yes, from the $20-a-month Plus plan, with usage limits growing on higher tiers. If you already pay for ChatGPT, trying Codex costs nothing more.
What are cloud agents?
Tasks running in isolated environments with a copy of your repository, in parallel and without occupying your machine. You launch them from the web or your phone and review the changes as proposals when they finish.
Codex or Claude Code?
If you already pay for one of the two subscriptions, start with the one you have: both agents are included. From zero, Claude Code holds the stronger reputation on long tasks and large codebases, and Codex answers with cloud parallelism. The full comparison develops the nuances.
Is the CLI open source?
Yes, the Codex CLI is open source and works with a ChatGPT subscription or on API consumption. The cloud side and the ecosystem integration are OpenAI’s product.
Best alternatives to OpenAI Codex
Guides that use OpenAI Codex
How to choose your coding agent: the decision in four questions
Guide to choosing a coding agent with judgment: what you already pay for, terminal or IDE, your…
How to work with several coding agents in parallel
Guide to parallel agent work: slicing independent tasks, orchestrating with Conductor or Codex's…