Our verdict · Coding agent
Codex is OpenAI's answer to Claude Code and, on published specs, an even match: a local agent in the terminal, IDE and desktop app, plus cloud tasks that run in parallel and GitHub code review. It comes with ChatGPT Plus at $20, and it publishes more about its limits than most rivals: per-model message ranges per five hours and credit rates per million tokens. It runs OpenAI models only, and like any coding agent it brings no hosting or database.
Pick it if
- Developers already paying for ChatGPT Plus, Pro or Business
- Teams that want cloud tasks and automatic GitHub PR review in one subscription
- People who want an open-source (Apache-2.0) CLI they can read and pin
Skip it if
- You want a hosted app without touching code or a terminal
- You want Claude or Gemini models in the same tool
- You need SSO without buying Business seats (Plus and Pro have no SAML SSO)
Codex is OpenAI's coding agent. It reads a repository, plans, edits files, runs commands and tests, and opens pull requests, from an open-source CLI, extensions for VS Code, Cursor and Windsurf, JetBrains IDEs, a desktop app, the web at chatgpt.com/codex, and iOS. Cloud tasks run in OpenAI's sandboxes while you do something else, and it can review pull requests on GitHub automatically. We haven't used Codex for client work the way we use Claude Code, so this review compares published plans and docs, opened on November 1, 2026. Mahmoud uses Claude Code regularly; read the comparison with Claude Code knowing that.
Scores
- Speed to a first version3/5
Fast once a repo exists, and cloud tasks start from the browser; still no hosting or canvas for a first demo.
- Depth before the wall5/5
Subagents, hooks, MCP servers, skills, AGENTS.md, sandbox and approval modes, and parallel cloud tasks.
- Backend and data2/5
Brings no backend; it works on whatever you choose.
- Ownership and exit5/5
Your repo and an open-source CLI; AGENTS.md instructions carry over to other agents.
- Pricing predictability4/5
Flat plan prices plus published per-model message ranges and credit rates per million tokens; ranges are wide and weekly limits aren't numbered.
- Team and production readiness4/5
Business adds SAML SSO and managed config at $20-25 a seat; RBAC, SCIM and audit logs need Enterprise.
Overall 3.8/5, the plain average of six criteria. Scored by Mahmoud Halat from published specs and docs, checked Nov 1, 2026. How we score.
Codex and Claude Code have the same shape, and on most criteria they score the same. Codex edges ahead on pricing predictability, for one reason: OpenAI publishes per-model message ranges for each five-hour window and credit rates per million tokens, where Anthropic publishes multiples of its Pro plan.
Pricing and limits
| Plan | Price (USD) | What you get |
|---|---|---|
| Free / Go | $0 / $8/mo | Desktop app only, with the smallest model at standard speed, subject to rollout |
| Plus | $20/mo | Web, CLI, IDE extension and iOS; cloud tasks, automatic code review and Slack; published per-model message ranges per 5 hours |
| Pro | $100, $200 or $500/mo | Everything in Plus with no five-hour limit currently; ultrafast mode on the $500 tier |
| Business | $20/user/mo annual ($25 monthly) | Two or more users; SAML SSO, managed config, no training on business data by default |
| Enterprise / Edu | Contract | RBAC, SCIM, audit logs, data residency, Codex Security; flexible credit-based usage |
| API key | API rates | CLI, SDK and IDE extension billed per token; no cloud tasks or GitHub review |
- Billed in
- ChatGPT subscription with per-model limits per 5 hours (weekly limits may apply); extra credits priced per 1M tokens
- Free tier
- Free and Go ($8) plans: the desktop app with a small model only, subject to rollout
Source: OpenAI: Codex pricing, openai/codex on GitHub. Prices checked Nov 1, 2026, before tax. Extra usage is bought as ChatGPT credits, priced per million tokens per model on the pricing page; fast mode costs 2 to 2.5 times as many credits. Usage figures are OpenAI's estimates.
Codex is priced as part of ChatGPT, so the question is less "what does it cost" than "how far does a plan go". OpenAI answers that more concretely than most:
- Plus ($20) lists local messages per five hours by model. The ranges are wide (5-45 on the largest model, 350-3,000 on the smallest) because a message's cost depends on how much code it reads and writes. Weekly limits may apply on top.
- Pro (from $100, with $200 and $500 tiers) currently has no five-hour limit. The $500 tier adds an "ultrafast" mode that uses credits at 6 to 8 times the normal rate.
- Business ($20 a user a month billed annually, or $25 monthly, two or more users) uses the same message estimates as Plus and adds SAML SSO, managed configuration and no training on business data by default. Plus and Pro don't have that default.
- Credits top up any plan. They're priced per million tokens per model, so a heavy month on the largest model costs many times a month on the smallest. OpenAI's own example puts a typical task on its mid-sized model at 2 to 15 credits.
- API key use bills per token at API rates and keeps the CLI, SDK and IDE extension, but drops cloud tasks and GitHub review. OpenAI recommends it for CI.
The practical choice for one developer is between Plus with occasional credits and Pro at $100. Watch the usage dashboard for two weeks on Plus: if you're buying credits most weeks, add up what they cost and compare it with the $80 step to Pro.
What it does well
Local and cloud from one subscription. You can run Codex in the terminal against your own checkout with a sandbox and approval mode that controls what it may touch, then hand a longer task to a cloud sandbox and pick up the pull request later. Both count against the same plan.
Configuration you can read. Project instructions live in AGENTS.md, the open file that several agents now read, so a repo set up for Codex also guides other tools. The config file supports MCP servers, lifecycle hooks, skills and subagent roles, and Business admins can lock settings with a managed requirements.toml.
Review on every pull request. Plus and above can review pull requests on GitHub automatically, and a local /review works on any plan. If your team's bottleneck is review rather than writing code, that matters as much as the agent.
An open-source CLI. The CLI is Apache-2.0 on GitHub, with 128,000 stars when we looked. You can read what it does, pin a version, and build it yourself.
Where it falls short
Codex runs OpenAI models only. If you want to switch to Claude or Gemini for a task, an editor like Cursor or Windsurf (now Devin Desktop) suits you better. The message ranges are honest but wide, so you still learn your real allowance from the dashboard, and weekly limits aren't numbered. Plus and Pro have no SSO, and the admin controls a larger company expects (RBAC, SCIM, audit logs) need Enterprise. Like every coding agent, it brings no hosting, database or visual editor: if you don't have a codebase yet, an app builder gets you to a demo faster.
Ownership
Everything stays in your repository, on your machine or in a cloud sandbox attached to a repo you choose. Instructions in AGENTS.md, MCP configuration and skills are plain files, so moving to another agent costs little.
Key facts
- Made by
- OpenAI
- Works in
- Terminal (open-source CLI), VS Code, Cursor, Windsurf and JetBrains IDEs, desktop app, web, iOS, GitHub (incl. Actions) and Slack
- Models
- OpenAI models only
- Backend
- None: it works on whatever stack your repo uses
- Ownership
- Your repository; the CLI is Apache-2.0 and reads AGENTS.md
Alternatives
- Claude CodeThe closest rival: same shape, Claude models, documented GitLab CI and scheduled tasks
- CursorIf you want an editor with several vendors' models
- GitHub CopilotIssue-to-PR work native to GitHub at a $10 seat
Head-to-head: Claude Code vs Codex
Questions people ask
Is OpenAI Codex free?
Partly. ChatGPT's Free and Go ($8) plans get the Codex desktop app with OpenAI's smallest model, subject to rollout. The CLI, IDE extension, web and cloud tasks start with ChatGPT Plus at $20 a month. You can also run the CLI and IDE extension on an API key, billed per token, without a subscription. Checked November 1, 2026.
What are Codex's usage limits?
On Plus, OpenAI publishes a range of local messages per five hours for each model, from 5-45 on its largest model to 350-3,000 on its smallest, and says weekly limits may also apply. Pro ($100, $200 or $500) currently has no five-hour limit. Business seats use the same estimates as Plus. OpenAI calls all of these estimates and points to the usage dashboard.
Is the Codex CLI open source?
Yes. The CLI is on GitHub at openai/codex under the Apache-2.0 licence, and installs with a shell script, npm or Homebrew. You sign in with a ChatGPT plan or an API key.
Is Codex the same as the old OpenAI Codex model?
No. The 2021 Codex was a code model behind the first GitHub Copilot and was retired. Today's Codex is an agent product, launched in 2025, that runs OpenAI's current models in the terminal, IDE, desktop app and cloud.