Skip to content

an open-source alternative to the dots, bots and muses of the world, run an office of claude code/codex like agents on your laptop, sandboxes or anywhere, uses your existing subscriptions

seo-engine: refresh codex-cli-vs-claude-code, review fixes

chaitanyagiri762b1778195f
diff --git a/blog/src/posts/codex-cli-vs-claude-code.md b/blog/src/posts/codex-cli-vs-claude-code.md
index 5790a515..11f13db1 100644
--- a/blog/src/posts/codex-cli-vs-claude-code.md
+++ b/blog/src/posts/codex-cli-vs-claude-code.md
@@ -11,20 +11,20 @@ secondaryKeywords: ["claude code vs codex", "codex cli vs claude code", "codex v
tags: ["Comparisons", "Claude Code", "Codex", "Multi-Agent", "Engines"]
faq:
- q: "Is Codex better than Claude Code?"
- a: "Neither wins outright as of 29 Sep 2026. Codex is the better fit if you want a free way in, limits published per model, and computer use that can run in the background on a Mac. Claude Code is the better fit if you want Opus 5.5 as the default model and computer use from inside the CLI on a Pro or Max plan."
+ a: "Neither wins outright as of 29 Sep 2026. Codex is the better fit if you want to try it free in the desktop app, limits published per model, and computer use that can run in the background on a Mac. Claude Code is the better fit if you want Opus 5.5 as the default model and computer use from inside the CLI on a Pro or Max plan."
- q: "Is Codex CLI free?"
- a: "Yes, with a cap. OpenAI's Codex pricing page lists a Free plan at $0 and a Go plan at $8 a month, both with GPT-6 Luna in the desktop app subject to rollout, checked 29 Sep 2026. The CLI, IDE extension and web start at Plus, $20 a month. Claude Code is not on Claude's Free plan at all; it starts at Pro."
+ a: "No. The CLI needs ChatGPT Plus ($20 a month) or higher, or an API key billed at API rates, per OpenAI's Codex pricing page checked 29 Sep 2026. The Free ($0) and Go ($8 a month) plans get GPT-6 Luna in the desktop app only, subject to rollout. Claude Code is not on Claude's Free plan either; it starts at Pro."
- q: "Which has higher usage limits, Codex or Claude Code?"
a: "You can only compare them on paper for Codex, because OpenAI publishes estimated messages per five hours for each model and Anthropic does not. Both products meter a rolling five hour window plus a weekly limit, and both sell credits once you run out. The multipliers line up: 5x and 20x tiers on each side."
- q: "Does Claude Code read AGENTS.md?"
- a: "Yes. Claude Code's memory docs, checked 29 Sep 2026, say it can read a repository's `AGENTS.md` on its own or alongside `CLAUDE.md`. Codex uses `AGENTS.md` natively, so one file can now brief both agents."
+ a: "Yes, with a catch. Claude Code's memory docs, checked 29 Sep 2026, say it reads `AGENTS.md` by default only when there is no `CLAUDE.md`, `.claude/CLAUDE.md` or `CLAUDE.local.md` in the working directory or above it. If one exists, it reads the `CLAUDE.md` files only, unless that file imports `AGENTS.md` or you change the Project instructions setting. It needs Claude Code v2.1.277 or later."
- q: "Does Codex CLI have computer use?"
a: "Not in the CLI. OpenAI's docs put Computer Use in the ChatGPT desktop app, with Codex, on macOS and Windows in supported regions, checked 29 Sep 2026. Claude Code has it inside the CLI, but only on macOS, as a research preview on Pro and Max plans."
- q: "Can I run Codex and Claude Code together?"
- a: "Yes. Munder Difflin, free and open source, runs Codex and Claude Code agents side by side on one office floor, each in its own terminal, and you choose the CLI per agent. Both tools also read `AGENTS.md` now, so one instructions file can cover both."
+ a: "Yes. Munder Difflin, free and open source, runs Codex and Claude Code agents side by side on one office floor, each in its own terminal, and you choose the CLI per agent."
---
-Codex and Claude Code are close enough in September 2026 that your plan should decide: pick Codex for a free tier, limits published per model and background computer use on a Mac; pick Claude Code for Opus 5.5 by default and computer use inside the terminal.
+Codex and Claude Code are close enough in September 2026 that your plan should decide: pick Codex for a free desktop tier, limits published per model and background computer use on a Mac; pick Claude Code for Opus 5.5 by default and computer use inside the terminal.
You can also skip the choice: switch between the two by hand, or run both in [Munder Difflin](https://harnessmd.com/download), free and open source, which puts Codex and Claude Code agents side by side on one floor, each in its own terminal. Everything below was checked on 29 Sep 2026 against each vendor's own pricing pages and docs.
@@ -34,22 +34,22 @@ You can also skip the choice: switch between the two by hand, or run both in [Mu
|:--|:--|:--|
| Where it runs | CLI, IDE extension, ChatGPT desktop app, web, iOS | CLI, IDE extensions, desktop app, web |
| Source | CLI is Apache-2.0 on GitHub | Closed, Anthropic commercial terms |
-| Models | GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, GPT-5.6 family. GPT-5.5 retires 14 Oct 2026 | Opus 5.5 is the default on Pro, Max, Team and Enterprise. Sonnet 5.5, Haiku, Fable 5.1 on usage credits |
+| Models | GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, GPT-5.6 family. The pricing page still lists GPT-5.5 (retires 14 Oct 2026), GPT-5.4 and GPT-5.4 mini | Opus 5.5 is the default on Pro, Max, Team and Enterprise, plus Sonnet and Haiku. Fable 5.1 can bill to usage credits, depending on plan and seat tier |
| Cheapest plan with the CLI | Plus, $20 a month (Free and Go get the desktop app only) | Pro, $20 a month |
-| Heavy plans | Pro 5x from $100, Pro 20x at $200 (closed to new sign ups since 10 Sep 2026) | Max 5x at $100, Max 20x at $200 |
+| Heavy plans | Pro 5x from $100, Pro 20x at $200 (new sign ups paused, reported 11 Sep 2026) | Max 5x at $100, Max 20x at $200 |
| Limits | Published ranges per model per 5 hours, weekly limits may apply, credits after | 5 hour session limit plus a weekly limit, no published message counts, credits after |
| Sandbox | On by default: Seatbelt on macOS, bubblewrap on Linux and WSL2, native Windows sandbox | Built in Bash sandbox (Seatbelt, bubblewrap), set up with `/sandbox`. No native Windows |
| Approval controls | `--sandbox` and `--ask-for-approval`, two flags | One `--permission-mode` with six modes |
| Computer use | ChatGPT desktop app on macOS and Windows, background on macOS, not in the CLI | In the CLI on macOS, research preview, Pro and Max only. Desktop app: macOS and Windows |
| MCP | Yes, STDIO and streamable HTTP | Yes, local stdio and remote HTTP |
| Subagents | On by default in current releases, plus custom agents | Built in and custom, each in its own context window |
-| Project instructions | `AGENTS.md` | `CLAUDE.md`, and it reads `AGENTS.md` too |
+| Project instructions | `AGENTS.md` | `CLAUDE.md`. Reads `AGENTS.md` only when no `CLAUDE.md` exists, by default |
-Sources: OpenAI's [Codex pricing page](https://learn.chatgpt.com/docs/pricing) and [Pro tiers article](https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers), Anthropic's [Max plan article](https://support.claude.com/en/articles/11049741-what-is-the-max-plan) and [Claude Code model docs](https://code.claude.com/docs/en/model-config), all checked 29 Sep 2026. Sandbox flags come from `codex --help` and `claude --help`, run the same day.
+Sources: OpenAI's [Codex pricing page](https://learn.chatgpt.com/docs/pricing), Anthropic's [Max plan article](https://support.claude.com/en/articles/11049741-what-is-the-max-plan) and Claude Code's model configuration docs, all checked 29 Sep 2026. Sandbox flags come from `codex --help` and `claude --help`, run the same day.
## Which one has higher usage limits, Codex or Claude Code?
-Only Codex lets you compare on paper, because OpenAI publishes estimated local messages per five hours for each model and Anthropic publishes none. OpenAI's [pricing page](https://learn.chatgpt.com/docs/pricing), checked 29 Sep 2026, lists these ranges:
+Only Codex lets you compare on paper, because OpenAI publishes estimated local messages per five hours for each model and Anthropic publishes none. OpenAI's pricing page, checked 29 Sep 2026, lists these ranges:
| Model | Plus | Pro 5x | Pro 20x |
|:--|:--|:--|:--|
@@ -59,11 +59,11 @@ Only Codex lets you compare on paper, because OpenAI publishes estimated local m
The same page says weekly limits may also apply, and that Plus and Pro users can buy credits when they run out. The spread inside each range comes from task size, context and reasoning, so a long Astra session lands near the bottom end.
-Anthropic describes its limits as multipliers instead. Max 5x gives five times Pro's per session usage and Max 20x gives twenty times, both with a weekly limit across all models that resets at a fixed time for your account, per its [Max plan article](https://support.claude.com/en/articles/11049741-what-is-the-max-plan) checked 29 Sep 2026. Anthropic also raised five hour limits on Pro, Max, Team and seat based Enterprise when it shipped [Opus 5.5 on 22 Sep 2026](https://www.anthropic.com/claude-opus-5-5). Claude Code and the Claude apps share one pool, so a long chat eats into your coding time. Our [Claude Code Max plan tips](/blog/claude-code-max-plan-tips/) cover how to stretch it, and the [Codex Max plan tips](/blog/codex-max-plan-tips/) do the same for OpenAI's side.
+Anthropic describes its limits as multipliers instead. Max 5x gives five times Pro's per session usage and Max 20x gives twenty times, both with a weekly limit across all models that resets at a fixed time for your account, per its Max plan article checked 29 Sep 2026. Anthropic's Opus 5.5 announcement on 22 Sep 2026 also raised five hour limits on Pro, Max, Team and seat based Enterprise. Claude Code and the Claude apps share one pool, so a long chat eats into your coding time. Our [Claude Code Max plan tips](/blog/claude-code-max-plan-tips/) cover how to stretch it, and the [Codex Max plan tips](/blog/codex-max-plan-tips/) do the same for OpenAI's side.
### Can new subscribers still get ChatGPT Pro 20x?
-Not as a new subscriber right now. OpenAI's [Pro tiers article](https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers) says that from 10 Sep 2026 it is temporarily pausing new sign ups and upgrades to Pro 20x, while existing Pro 20x subscriptions and all Pro 5x plans carry on. If you burn through Pro 5x, your options today are credits, an API key, or a second tool. Anthropic's Max plan article carried no pause of that kind on 29 Sep 2026.
+Probably not right now. [Fortune reported on 11 Sep 2026](https://fortune.com/2026/09/11/openai-astra-chatgpt-pro-pause/) that OpenAI had paused new sign ups and upgrades to the $200 ChatGPT Pro tier because of demand for GPT-6 Astra, with existing Pro accounts unaffected. We could not load OpenAI's own help page to confirm it, so check before you plan around it. If you burn through Pro 5x, your options are credits, an API key, or a second tool. Anthropic's Max plan article carried no pause of that kind on 29 Sep 2026.
## Does Codex or Claude Code have better computer use?
@@ -95,18 +95,18 @@ Codex applies its sandbox in the default permissions mode, so every command it s
## Does Claude Code read `AGENTS.md`?
-Yes, and that changed since we first wrote this post. Claude Code's [memory docs](https://code.claude.com/docs/en/memory), checked 29 Sep 2026, say it can read a repository's `AGENTS.md` on its own or alongside `CLAUDE.md`. Codex uses `AGENTS.md` natively and creates one with `/init`. A repo that runs both no longer needs two copies of the same rules or a symlink between them.
+Yes, but only when there is no `CLAUDE.md`, by default. Claude Code's memory docs, checked 29 Sep 2026, say it reads `AGENTS.md` when there is no `CLAUDE.md`, `.claude/CLAUDE.md` or `CLAUDE.local.md` in your working directory or above it. If one of those exists, it reads your `CLAUDE.md` files only. To share one file, import `AGENTS.md` from `CLAUDE.md`, or change the Project instructions setting to read both. Direct reading needs Claude Code v2.1.277 or later. Codex uses `AGENTS.md` natively and generates one with `/init`, so a repo with only an `AGENTS.md` now briefs both agents.
## Is Codex CLI free?
-Not quite: the free tier covers the desktop app, not the CLI. OpenAI's pricing page, checked 29 Sep 2026, lists Free at $0 and Go at $8 a month, both with GPT-6 Luna in the desktop app subject to rollout, and Plus at $20 a month for Codex on the web, in the CLI, in the IDE extension and on iOS. You can also run the CLI on an API key and pay API rates. Claude Code is not on Claude's Free plan; it starts at Pro.
+No: the free tier covers the desktop app, not the CLI. OpenAI's pricing page, checked 29 Sep 2026, lists Free at $0 and Go at $8 a month, both with GPT-6 Luna in the desktop app subject to rollout, and Plus at $20 a month for Codex on the web, in the CLI, in the IDE extension and on iOS. You can also run the CLI on an API key and pay API rates. Claude Code is not on Claude's Free plan; it starts at Pro.
{% img "note-2" %}
## Pick Codex if..., pick Claude Code if...
**Pick Codex if:**
-* You want to start free before you commit.
+* You want to try Codex free in the desktop app before you pay.
* You like knowing roughly how many messages each model buys you per five hours.
* You want computer use that runs in the background on a Mac or works on Windows.
* You want to read the CLI's source (Apache-2.0).
@@ -121,4 +121,4 @@ If you are still choosing between whole categories of tools, not just these two
## Can you run Codex and Claude Code together?
-Yes, and for a lot of teams that is the honest answer. Munder Difflin (as of 0.5.3) lets you pick the CLI per agent, Claude Code or Codex among others, and runs each one as a real terminal process on the same floor. A Codex agent can take the UI bug while a Claude Code agent takes the refactor, and both read the same `AGENTS.md`. When one plan hits its weekly limit, the other keeps working.
+Yes, and for a lot of teams that is the honest answer. Munder Difflin (as of 0.5.3) lets you pick the CLI per agent, Claude Code or Codex among others, and runs each one as a real terminal process on the same floor. A Codex agent can take the UI bug while a Claude Code agent takes the refactor. When one plan hits its weekly limit, the other keeps working.