Claude Code or Codex: which to choose

1. At a glance

Claude CodeCodex

Entry price

from $17/mo

Pro with annual billing

✓from $0

Included in all ChatGPT plans

Free plan

No

The Free plan does not provide access

✓Yes

In the ChatGPT Free plan

Context window

✓1M

Right away on any paid plan

up to 1M

GPT-6's window, switched on in Codex with a setting

Open source

No

Proprietary CLI

✓Yes

CLI Apache-2.0, ~126K stars

Surfaces

✓14

CLI · IDE · desktop · web/iOS · CI · Slack

11

app · CLI · IDE · cloud · GitHub · mobile

No

No

Not in supported countries

2. These are different form factors, and that is the key point

Claude Code was born as a terminal agent and still lives primarily in the terminal: it grows into an existing repository, reads CLAUDE.md, keeps project memory, runs tests and assembles pull requests. Around the core, 14 surfaces have grown: CLI, extensions for VS Code and JetBrains, a desktop app, web and iOS, CI integrations, Slack, SDK. But the character is the same: it is a tool for serious work in existing code, not for autocomplete.

Claude Code · terminal

Codex is built the other way around: it is one product with several windows into it. It lives in the ChatGPT app, in a desktop app with parallel threads and worktrees, in the CLI, in an IDE extension, in the cloud, in GitHub reviews and even on your phone. Notable features: computer use on macOS and Windows, cloud tasks with best-of-N, scheduled automations and Sites, website deployment on OpenAI hosting.

Codex · app

The practical takeaway: if your work is one large repository and deep changes within it, Claude Code is a better fit. If you want a pipeline of parallel tasks of different kinds from any device, Codex is a better fit.

3. Only Codex lets you start for free

Codex is included in every ChatGPT plan, including Free at $0. On Free and Go at $8 it runs in the desktop app on the fast GPT-6 Luna; the higher GPT-6 Sol and Astra unlock with Plus at $20. Above that is Pro from $100 a month, and new subscriptions to the $200 plan have been paused since Sep 10.

Claude Code has no free entry point: the Free plan gives no access. Entry costs $17 a month for Pro billed annually, or $20 month to month; heavy scenarios live on Max at $100-200.

The limits work in a similar way: a 5-hour window with a weekly cap on top of it. In Claude Code the pool is shared across chat, desktop and terminal; in Codex it is split between local work and cloud tasks.

Limits are a living system, not a constant from a price list. In July OpenAI removed the 5-hour window in Codex on paid plans, and on Aug 25 brought it back on Plus. On Pro there is no window yet, only the weekly cap.

4. Numbers duel: look at the AA Index, not promises

Comparing agents in words is pointless, that is what the AA Coding Agent Indexis for: cost and time on real tasks, with a separate score for each agent plus model pair. Latest snapshot as of Nov 7, 2026:

Here is how to read it: the score belongs to the pair, not to the agent in general. Switching the model inside an agent moves the result more than switching the agent itself, and the leader flips with major model releases. The full ranking and methodology are open: AI agent ranking.OpenCode for teams is evaluated on model, execution permissions and repository handling.

The score is not the only number. In the Sep 25 snapshot an average task on Claude Code with Opus 5.5 cost $13.04 and took 65 minutes. Codex with GPT-6 Astra did it for $7.47 in 29 minutes, almost twice as cheap and twice as fast with a score four points lower.

5. Models and context: the windows have evened out

Since Sep 22, Claude Code runs on Opus 5.5 by default, and since Sep 28 there's Sonnet 5.5 for quick tasks. The heaviest, Fable 5.1, is paid for on Pro with credits on top of the subscription. All three have a 1M token window on any paid plan: it fits a sizable part of an average repository.

Codex got GPT-6 Sol and GPT-6 Luna on Sep 22, and via the API they cost half as much as the previous generation. OpenAI's top model, GPT-6 Astra, has been in Codex since Sep 4. GPT-6 models have a 1.05M-token window, but in Codex its size is set in the config, while Claude Code has 1M enabled out of the box.OpenAI Codex for teams is tested on a task with limited permissions and clear acceptance criteria.

On the other hand, Codex is open source: the CLI is on GitHub under Apache-2.0, written in Rust, with about 126,000 stars. For teams that care about reading and patching their tools, this is a real advantage: the Claude Code CLI is proprietary.Working with Aider is accepted based on the git change and the check results.

6.  

7. What we build this site with

Agents vibecode vibecoding.tech; a human writes the rules and accepts the work. As of September 8, 2026, nine-tenths of the site was made in Claude Code and one-tenth in Codex: of 276 merged changes, 250 were by the former and 26 by the latter.

The roles split according to each one's strengths. Claude Code builds pages, data, and checks; Codex reviews news before publication and generates images. What is happening right now is shown by live machine feed.

8. Which to choose

What changed in recent months