Codex vs Claude Code: which to pick, and how I run both

I run Claude Code and OpenAI's Codex on the same machine. Claude Code is where I work. Codex takes batches of implementation that Claude Code hands it. In October 2026 my machine logged 1,051 Codex sessions, and 919 of them were non-interactive codex exec runs, the kind a script or another agent starts.

Below is what each costs today, how they differ, and how to tell whether you need one of them or both.

Claude Code vs Codex, the short answer

Pick Claude Code if you want one agent you work with all day in the terminal: it reads the codebase, plans, edits and reviews.
Pick Codex if you pay for ChatGPT Plus or Pro, since Codex comes with it, or if you want to spend nothing to start. Codex has a free tier with limited usage. Claude Code has none.
Run both if your work comes in batches of independent tasks. Plan in one, hand the batch to the other, and review the diff. That is my setup, described below.

What each costs

Prices as the vendors' pages showed them: Anthropic's on November 4, 2026, and OpenAI's on November 5, 2026.

Claude Code (claude.com/pricing):

Free: no Claude Code.
Pro: $20 a month, or $17 a month billed annually ($200 up front). Claude Code included. Best for: trying Claude Code on real work at the lowest price that includes it.
Max: $100 a month, with five times Pro's usage every five hours and every week. Best for: running into Pro's limits.
Ultra: $200 a month, with ten times Pro's weekly usage and, for "verified accounts only" in Anthropic's words, no five-hour limit. Best for: running into Max's limits.

Codex (ChatGPT pricing and Codex pricing):

Free: $0, limited Codex. Best for: trying Codex before you pay.
Go: $8 a month, limited Codex. Best for: a cheaper ChatGPT plan, if Codex is not why you pay.
Plus: $20 a month. Codex on the web, in the CLI, in the IDE extension and on iOS, plus cloud features such as automatic code review and a Slack integration. Best for: using Codex for real work at the entry price.
Pro: $100, $200 or $500 a month. OpenAI says Pro plans currently have no five-hour limit, and the $500 plan adds Ultrafast mode. Best for: running into Plus limits, or work that would stop at the five-hour window.
API key: Codex in the CLI, SDK or IDE extension, billed at API prices, without the cloud features. Best for: scripts and CI billed by usage instead of a monthly plan.

Neither vendor promises a number of tasks. OpenAI's own estimate for Plus is 15 to 160 local messages every five hours on GPT-6.1 Sol, depending on the task, and its pricing page says not to use API token prices to estimate what a plan includes.

From nothing up: Codex Free at $0 and ChatGPT Go at $8, both with limited Codex; $20 for one tool, either ChatGPT Plus (OpenAI lists its Codex as expanded) or Claude Pro (the cheapest plan with Claude Code); $40 for both.

If you hit the five-hour limit every day, two plans lift it: ChatGPT Pro, from $100, which OpenAI says currently has none, and Claude Ultra at $200, for verified accounts.

How I split the work

Claude Code plans, reads the codebase and writes the small changes. When a plan has five or more separate pieces of implementation, Claude Code hands that batch to Codex with an explicit instruction to test. Codex edits, and Claude Code reviews the diff. The rule is written down in hov-stack, the agent setup I publish on GitHub: the Codex hand-off rule.

Most of my October Codex sessions were codex exec runs, which is how a script or an agent starts Codex: it takes a task and a working directory (--cd) and exits when it is done. Claude Code has the same mode, claude -p, so either one can be the tool that hands off.

My Codex default is GPT-6.1 Sol, the model OpenAI's Plus estimate above is quoted for, at low reasoning effort, the Codex setting for how much the model reasons before it answers. It ran 464 of the 1,051 October sessions, and most of the rest ran GPT-6 Luna, GPT-6 Astra or GPT-5.6 Luna. Comparisons that tested GPT-5.5 are out of date: OpenAI retired it from Codex with ChatGPT sign-in on October 14, 2026.

How they differ

Working style. Both run as an interactive session in a terminal. Codex on Plus and Pro also runs on the web and adds cloud features, automatic code review and a Slack integration, which suit work you start and check on later.
Hand-off and scripts. codex exec and claude -p both run a task without an interactive session. That is what makes running both practical.
Instruction files. Codex reads AGENTS.md. Claude Code reads CLAUDE.md. I keep one file: AGENTS.md holds the instructions, and CLAUDE.md is a single line, @AGENTS.md, which imports it. Both tools then follow the same rules, so moving between them means no rewrite.
Limits. Both meter usage in five-hour windows and weekly limits on the lower plans. At the top, ChatGPT Pro currently has no five-hour limit and Claude Ultra has none for verified accounts. Claude Max still has one.
Your data. OpenAI's pricing table says content on Free, Go, Plus and Pro is used to train its models, with an opt-out available. Anthropic's privacy center says chats and Claude Code sessions on its consumer plans, such as Free, Pro and Max, are used to improve its models if you allow it in your privacy settings or opt in another way, such as its Trusted Tester Program. Conversations flagged for safety review may also be used to train the models its Safeguards team uses, and Incognito chats are not used.

Which one should you start with

You want one agent to work with all day: Claude Code, from Pro at $20 a month.
You already pay for ChatGPT: start with Codex. Plus and Pro include it with expanded usage; on Go it is limited.
You want to spend nothing: Codex's free tier, which is limited. Claude Code starts at Pro, $20 a month.
You hit limits every day: ChatGPT Pro, which currently has no five-hour limit, or Claude Ultra, which has none for verified accounts.
Your work comes in batches: run both. Plan and review in one, implement in the other.

What changed

November 5, 2026: first version. Prices read November 4 (Anthropic) and November 5 (OpenAI). Versions on my machine: Claude Code 2.1.318 and codex-cli 0.166.0.

More on what I build is on the about page.