Comparisons

Codex vs Claude Code: What Each Costs, Limits and Lets the Agent Do

Published Updated 4 min readBy Vibe Coding Intel editorial

Compiled from official docs and public sources, checked .

The short answer: both are terminal-and-editor coding agents sold through a chat subscription, and at the entry paid price they are level: Codex on ChatGPT Plus and Claude Code on Claude Pro are each $20 a month. The difference that matters for budgeting is disclosure. OpenAI publishes a usage range for Codex (for example 15 to 150 messages per five hours on Plus, depending on the model); Anthropic publishes only multiples for Claude Code. The difference that matters for safety is the default: Codex starts offline inside a workspace sandbox, while Claude Code (on current versions) starts in auto mode, where a second model reviews actions, and asks you only when its reviewer flags something.

The price ladder side by side

Monthly prices from OpenAI’s Codex pricing page and claude.com/pricing, checked 29 September 2026. The Claude Max tiers come from the figures on our Claude Code pricing page.

RungCodexClaude Code
Free$0, basic capabilitiesNone (Pro required)
LightGo, $8None
StandardPlus, $20Pro, $20 ($17 a month billed annually)
HeavyPro from $100 (5x limits), or 20xMax $100 (5x) or $200 (20x)
TeamBusiness $25 monthly, $20 per user on annual billingTeam standard seat $25 monthly, $20 billed annually
Pay per tokenAPI key at standard API ratesAPI key at standard API rates

The ladders are close to a mirror image, with Codex adding a cheaper $8 tier and a free one. One trap on both: if an API key is set in your environment, the tool can bill you per token instead of drawing on the subscription.

What each vendor says about allowance size

Plan at about $20Stated allowance
Codex Plus5 to 45 messages per five hours on the largest model, 15 to 150 on the next, 350 to 3,000 on the smallest
Claude ProFive-hour and weekly limits shared with Claude chat; size not published

The Codex ranges are wide, an order of magnitude from floor to ceiling, so they set expectations rather than promise a number, but they are still more than Claude states. Past the included limit, Codex sells credits at listed rates per million tokens; Claude sells usage credits billed at API rates. Anthropic’s docs offer a real-usage anchor instead: about $13 per developer per active day across enterprise deployments, per its cost guide.

Default posture: who asks first

AttributeCodexClaude Code
Starting behaviourWorkspace-write with approval on request: edits and commands inside your workspace without asking; asks before leaving it or using the network. Folders not under version control start read-onlyAuto mode from v2.1.283 (a second model reviews actions); earlier versions start in auto only on Pro, Max and Team, otherwise Manual, where edits and shell commands prompt
NetworkOff by default; must be enabled in settingsSandboxed commands get no network access by default; other network actions go through permission checks
IsolationOS-enforced sandbox limited to the workspacePermission rules and modes, with optional sandboxing
Cloud runsSetup phase online, agent phase offline unless enabledNot compared here

Both let you change the posture, and Claude Code’s Manual mode (default) is one Shift+Tab away (permission modes). So the defaults differ in mechanism: Codex relies on a fence around the workspace, Claude Code on a reviewing model plus permission rules. Which suits you depends on whether you would rather trust the boundary or the reviewer.

Which to pick

  • Need a cheap or free way in: Codex has the $8 and free tiers.
  • Want to know the limit before you buy: Codex publishes a range.
  • Already pay for Claude chat: Claude Code adds no price, but shares the allowance with chat.
  • Prefer to see every edit first: Claude Code in Manual mode.
  • Wider field: the best AI code editor guide puts both against IDE assistants, Claude Code vs Cursor and Claude Code vs GitHub Copilot cover the other pairings, and the Claude Code review covers the tool itself.

We do not rank the two on code quality: that would need benchmark evidence we have not compiled, and model names on both sides change every few months.

Method

Researched from OpenAI’s Codex pricing and security documentation and Anthropic’s pricing and Claude Code documentation on 29 September 2026. We have not run either tool and measured nothing. OpenAI’s pages redirected from developers.openai.com to learn.chatgpt.com during the check, so the links above use the current address. Re-checked every 30 days.

Common questions

Is Codex cheaper than Claude Code?

At the entry paid tier they are equal at $20. Codex adds an $8 tier and a free tier; Claude Code has neither.

Do Codex and Claude Code share one allowance with chat?

Anthropic states that Pro and Max usage is shared between Claude chat and Claude Code. The Codex page lists Codex limits by plan and does not say how they interact with chat.

Which is safer by default?

Codex blocks network access and fences edits to the workspace; Claude Code’s default auto mode has a second model review actions, and Manual mode asks before edits and commands. Different mechanisms, and both are adjustable.

Sources

Update log

  1. New page. Codex plans and credits read on OpenAI's Codex pricing page; approval defaults from its agent security docs; Claude Code from Anthropic's pricing and permission docs.