Claude Code vs Codex: Which Agent Fits?

The Short Answer

Claude Code and Codex are the two coding agents most developers compare first. Both read a repository, edit files, run commands and tests, and hand back a reviewable diff. Both run in a terminal, inside common editors, in a desktop app and in the cloud. The differences that decide a purchase are the models behind each agent, which subscription pays for the work, and how each product handles review, permissions and team administration.

Choose Claude Code when you want Anthropic's models, led by Claude Opus 5.5 and Sonnet 5.5, driving a shell-first workflow, or when your company already buys Claude plans or routes models through Amazon Bedrock, Google Cloud or Microsoft Foundry. Choose Codex when your team already pays for ChatGPT, wants OpenAI's GPT-6.1 Sol and GPT-6 Astra models, or wants cloud tasks and GitHub pull-request review included from the $20 Plus plan.

Prices and plan details in this guide were checked on the official Anthropic and OpenAI pricing pages and product documentation on September 30, 2026. Both vendors ship model and plan changes often, so confirm the current rollout for your account before you commit. Claude pricing, Codex pricing.

How Each Agent Works

Claude Code is a command-line agent with extensions for VS Code and JetBrains, a desktop app and cloud sessions at claude.ai/code. In Claude Code, the opus alias resolves to Claude Opus 5.5 and the sonnet alias to Claude Sonnet 5.5 on the Anthropic API, and Fable 5.1 is available for the longest tasks. Effort levels run from low to max, with medium as the default for Opus 5.5 and Sonnet 5.5. Permissions use allow, ask and deny rules, and auto memory carries useful notes between sessions. Claude Code model configuration, Claude Code permissions.

Codex is one agent that runs in the ChatGPT desktop app, the Codex CLI, an IDE extension, on the web and on iOS, with Codex Cloud for delegated tasks. GPT-6.1 Sol, released on September 29, 2026, is the workhorse model on Plus and Pro, GPT-6 Sol remains available, GPT-6 Astra handles the hardest reasoning, and GPT-6 Luna covers light tasks. Codex uses sandboxing and permission controls, and repository guidance lives in AGENTS.md files. Codex models, Codex permissions.

Both agents support parallel work in Git worktrees, subagents, MCP servers, skills and plugins. Worktrees isolate checkouts but not shared services such as databases, so parallel agents still need clean test environments.

Where Claude Code Wins

Anthropic's models in a shell-first agent. Claude Code is built around the Claude models, and its model aliases, effort control and 1M-token context windows are tuned for them. Opus 5.5 costs $4 input and $20 output per million tokens on the API, with cache hits at $0.20, and Sonnet 5.5 costs $2/$10. For teams that prefer Claude's style on long refactors, that is the most direct route. Claude API pricing.

Cloud procurement flexibility. Claude Code documents routes through Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry, so a company can keep inference inside an existing cloud contract. Model aliases can resolve differently on those providers, so confirm the exact model before rollout. Codex also supports OpenAI models through Amazon Bedrock, so AWS alone does not settle the choice. Claude Code platforms, Codex on Amazon Bedrock.

Seat types for heavy users. Claude Team lets an organization mix Standard seats with Premium seats that carry five times the usage. That suits a team where a few engineers run agents all day and most use them occasionally.

Where Codex Wins

Cloud tasks and review from the $20 plan. ChatGPT Plus includes Codex on the web, in the CLI, in the IDE extension and on iOS, plus cloud integrations such as automatic GitHub code review and Slack. Claude's managed Code Review is limited to Team and Enterprise as a research preview, although local diff review works on every Claude plan. Claude Code Review.

Lower entry points. Codex has limited access on the Free plan and on Go at $8 a month, both using GPT-6 Luna in the desktop app. Claude Code starts at Claude Pro, and the free Claude plan does not include it.

A cheap workhorse model. GPT-6.1 Sol costs $2/$10 per million tokens with cached input at $0.10, and Codex credits bill it at 50 input, 2.5 cached and 250 output credits per million tokens. OpenAI's usage estimates give Plus about 15 to 160 local GPT-6.1 Sol messages per five hours, against 5 to 45 on GPT-6 Astra, so choosing Sol for routine work stretches the same allowance much further. OpenAI API pricing.

Pricing and Usage

Both subscriptions share one usage pool across chat and coding. Claude's plans reset limits on a rolling five-hour window with weekly limits on top, and Claude Code draws from the same pool as Claude chat. ChatGPT Work and Codex share usage in the same way, and Pro plans currently have no five-hour limit. When you reach a limit, Claude lets paid users turn on usage credits billed at standard API rates, and ChatGPT Plus and Pro users can buy credits.

Claude Max comes in 5x and 20x versions of Pro usage, from $100 a month. Fable 5.1 on Claude plans works differently: Pro pays for it with usage credits, while Max can spend up to half of its weekly limits on Fable models. Fable plan access.

Authentication decides who pays. In Claude Code, an ANTHROPIC_API_KEY in the environment takes precedence over a subscription login and bills the API. Codex with an OpenAI API key bills at API prices and loses cloud features such as GitHub review. Run /status in either CLI before a long unattended job. Claude subscription authentication, Codex authentication.

Speed costs extra on both. Claude Code's Fast mode runs Opus 5.5 up to 2.5 times faster at $8/$40 per million tokens, is paid only from usage credits on subscription plans, and charges the whole existing context at the Fast input rate the first time you turn it on in a conversation. Codex's Fast mode uses included subscription usage at 2.5 times the Standard rate. Claude Code Fast mode.

Workflow and Safety Differences

Memory works differently. Claude Code's auto memory is on by default and saves useful notes across sessions. Codex offers memories as well, and GPT-6 Astra adds an experimental context-management mode in Codex that keeps notes across context windows within a task. Neither guarantees perfect recall, so durable project rules still belong in CLAUDE.md or AGENTS.md files that you review like code. Claude Code memory.

Desktop control is also bounded. Claude's computer use in the desktop app is an opt-in research preview for Pro and Max on macOS and Windows. Codex's Computer Use, browser control and Record & Replay are limited to specific regions. If your tests depend on driving a browser or native app, confirm that the feature works in your region and plan before building a workflow around it.

Review is where both products still need a human. Each agent can review its own diff or a pull request, but approving a permission prompt only allows an action; it does not show that the change is correct. Keep required tests and human approval on the merge step, whichever agent you choose.

Decision Framework

How to Run a Fair Trial

Test both agents on the same repository revision with three tasks: a bounded bug fix, a multi-file change and a longer refactor. Record the model, effort level and speed mode, the time to an acceptable diff, the number of retries, the tests that pass, and the usage or charges shown in each dashboard. Judge the result by cost per accepted change, not by cost per message.

Check permissions at every boundary: files, shell commands, network access, connected services and desktop control. Approving an action lets the agent act; it does not prove the change is correct, so inspect the diff and the test evidence before merging.

If both agents meet your bar, keep the one your organization already pays for. Switch only when the other one repeatedly saves enough review and rework time to justify a second subscription.