Codex vs Claude Code 2026: how to choose a tool

Learn more about working with models in the Ailoot documentation.

How do Codex and Claude Code differ?#

In April 2025, OpenAI released a new product, Codex CLI, together with the o3 and o4-mini models. That was the start of Codex as an open-source tool—first a wrapper around o3/o4-mini, then a standalone model, GPT-5-Codex, and by 2026, an entire lineup: GPT-5.3-Codex, GPT-5.4, GPT-5.5, with a reasoning effort switch from minimal to xhigh.

Anthropic released Claude Code earlier, in the summer of 2025, as part of its Claude Pro subscription. By 2026, it had become a full ecosystem: a VS Code extension, a web version, mobile notifications, a plugin marketplace, and managed agents with memory between sessions.

Technically, both are agents with the same basic capabilities:

• read and edit files in your project,
• run commands in a terminal,
• control a browser (Computer Use),
• connect MCP servers (Model Context Protocol) to work with external systems,
• store project instructions in a special file (CLAUDE.md for Anthropic, AGENTS.md for Codex),
• launch sub-agents to work in parallel.

The differences begin at the level of philosophy. By default, Claude Code asks permission before making changes, explains the pros and cons of a decision, and shows its reasoning. Codex by default works in an isolated sandbox and pushes through to the end of the task, putting anything it couldn't do right away in TODO. See the description of isolation modes in the official OpenAI documentation.

You can even hear this in the interface: with Claude Code, you hear Claude as a conversation partner. With Codex, you hear an executor.

Why the choice is not where people look#

Before comparing Codex and Claude Code by their capabilities, we need to take a step back. In the scenario described, almost all the cloud IDEs for vibe coding that are getting attention right now were tried: Bolt, Lovable, Replit Agent, v0, and Emergent. Each time, the conclusion was the same.

A cloud IDE is a sandbox. A good sandbox, convenient and beautiful. But a sandbox. You can't see your code. You don't control the architecture. You burn through tokens 10–20 times faster than on a proper stack because they run through their own billing on top of Anthropic or OpenAI. And worst of all, you can't take the project with you if you decide to leave.

The author describes costs comparable to three months of his Claude Code subscription; he still has enough resources in the subscription for two more projects.

This is a typical path for an entrepreneur in vibe coding:

1. Stage 1: sandbox

   Went into Bolt or Lovable, put together a prototype in an evening. Excitement.

2. Stage 2: tried to scale

   I wanted to turn the prototype into a production project. I needed control over the code, my own database, my own authentication, proper Git. The sandbox doesn't allow it.

3. Stage 3: hit the ceiling

   Burned through the budget 10 times faster than planned. Moving the project to my own stack is no longer possible—all the templates, databases, and files are locked inside the service.

4. Stage 4: moved to a proper stack

   Opened VS Code, installed Claude Code or Codex, spent a couple of days learning it—and felt like I was finally working like a professional.

The same logic applies to Cursor. Cursor is a copy of VS Code with AI capabilities bolted on. Google's Antigravity is a copy of Cursor with multi-agent features bolted on. The further a fork gets from the original, the more it locks you into someone else's ecosystem.

The real choice looks like this:

1. Cloud IDE (Bolt, Lovable, Replit Agent, v0): convenient for prototyping, but ties you to the platform. Suitable for testing an idea over a weekend.
2. Semi-IDE (Cursor, Antigravity, Trae): a development environment, but a VS Code fork that ties you to someone else's ecosystem. Cursor works through its own billing on top of provider APIs—that is, the Claude and GPT you connect to run through Cursor's tokens, not directly.
3. Open stack (VS Code + Codex / VS Code + Claude Code): you keep the familiar editor, native GitHub, file tree, and extensions, and get a full-fledged agent on top.

If you're at the third level, the difference between Codex and Claude Code becomes tactical. You can even work with both at once: Codex in one worktree, Claude Code in another; git worktree separates them into different branches.

And that's where the second rule comes in: a tool gives you speed, but without a methodology, speed works against you. The latest Opus 4.7 in the hands of someone who hasn't refined the idea, laid out the architecture, or approached the project through Plan Mode and CLAUDE.md produces garbage, just faster than usual. 99% of vibe-coding projects are abandoned in their early stages precisely because the idea wasn't fully worked through.

What Codex and Claude Code really cost in 2026#

The first thing people ask is how much a month will cost. Here are the figures, with caveats:

Sources for verification: Codex pricing and Claude pricing.

What you need to understand about limits:

• 5-hour windows are a small “basket” of messages that resets every five hours. In a window, Plus users can send 15 to 80 local messages with GPT-5.5 on Codex. On Pro 20x, it's 300 to 1,600. Exact figures depend on which models you call.
• Weekly cap is a weekly ceiling on top of the 5h windows. In May 2026, Anthropic raised it by 50% through July 13 for all Pro, Max, Team, and seat-based Enterprise users. For Max 20x, that works out to about 300 hours of Opus 4.7 per week instead of the previous 200.

If you do a lot of automation through the API (for example, run an agent in the background on a server), the actual bill can easily exceed $500 a month. This is about heavy API usage, not ordinary CLI use.

Plan recommendation for your situation:

• Fewer than 30 requests a day, ordinary edits: Plus or Pro $20 for either one should be more than enough.
• 30–100 requests a day, long agent sessions: Codex Pro $100 or Claude Max 5x $100.
• More than 100 requests a day, multi-hour autonomous tasks: Claude Max 20x $200 (with the May boost of +50%, it gives you a steady window) or Codex Pro $200 (more powerful for parallel runs and Computer Use).

The same budget in Bolt or Replit Agent will disappear an order of magnitude faster.

Who thinks longer: reasoning levels in Codex and Claude Code#

Reasoning effort is a lever for controlling how deeply the agent considers a task. At minimal, Codex solves it like a simple snippet generator: quickly and cheaply, without extended reasoning. At xhigh, it spends several times more tokens, lays out a tree of options, checks edge cases, and verifies its own work.

Codex levels (source: the official prompting guide):

Switch with the /effort high command in the TUI without restarting, or add the line model_reasoning_effort = "high" to ~/.codex/config.toml.

Claude Code has the same set of levels, but switches them using the interactive /effort slider, introduced in April 2026 (Week 16 changelog). Plus, since 2.1.x, hooks can see the effort.level value and $CLAUDE_EFFORT variable, so you can automatically invoke a different hook for a high reasoning task.

Practical advice: don't try to “run everything at max.” The strict rule is medium by default, with a manual switch to xhigh only when a task is genuinely complex (a new feature, architecture, a major refactor). For more on saving tokens, see the guide to Claude Code limits.

CLAUDE.md vs. AGENTS.md: how they differ and why you need both#

These are two files that work as “instructions for an agent” in your project. Both are ordinary Markdown and both can live in the repository root or in subfolders. The difference is which tool reads them.

CLAUDE.md is Anthropic's format, exclusive to Claude Code:

• Supports @import syntax: @AGENTS.md, @.claude/rules.md.
• Understands path-scoped rules (a CLAUDE.md in a subfolder applies only to code in that subfolder).
• Knows about permission levels, sub-agent delegation, and sessions.
• Sub-agent delegation and hooks are configured through references in CLAUDE.md.

AGENTS.md is an open standard announced at agents.md:

Codex reads the AGENTS.md hierarchy from the global ~/.codex/AGENTS.md to the repository root and then in each subfolder. The nearest file takes priority. Amp, Google's Jules, Cursor, and Factory also support the same format, so one file becomes universal context for any agent in the project.

⚠️ Important footnote: starting with version 2.1.277 (released September 18, 2026), Claude Code reads AGENTS.md on its own, but only if there is no CLAUDE.md, .claude/CLAUDE.md, or CLAUDE.local.md in the working folder or above it. If CLAUDE.md exists, Claude Code reads only that by default, and AGENTS.md is included with the line @AGENTS.md inside CLAUDE.md or a symlink ln -s AGENTS.md CLAUDE.md. The line @AGENTS.md does not make it read twice. You can change this behavior in /config, under Project instructions—for example, to make Claude Code read both files at once. All options, step by step, are in the guide to using AGENTS.md and CLAUDE.md in one project.

If you work in both tools and need rules just for Claude, the best strategy is:

# CLAUDE.md
@AGENTS.md

## Claude-specific
- Use sub-agent for security review
- Permission level: workspace-write

That is, one canonical AGENTS.md plus a thin layer of Claude-specific rules in CLAUDE.md that imports AGENTS.md with @.

What carries over 1-to-1 between the files:

• Commands (build, test, lint).
• Code style rules.
• Architecture description.
• Prohibitions on using certain patterns.
• Business context.

What remains only in CLAUDE.md:

• Sub-agent delegation instructions.
• @import chains.
• Path-scoped rules (CLAUDE.md in subfolders with different rules).
• Permission levels.
• Output Style preferences (discussed below).

This is the first part of the project's second brain: files that provide context to the agent.

What both released in May 2026#

Our “recent overview” from a month ago is already outdated. The May 2026 changelog review for both tools collects the changes—here's what matters.

Anthropic Claude Code (according to code.claude.com/docs/en/whats-new):

• Week 20 (May 11–15): Agent View—the claude agents command shows a single screen for all current Claude Code sessions: what's running, what's blocked, and what's done. /goal—the agent keeps working across turns until it reaches the specified condition. Fast mode switched to Opus 4.7 by default. Rewind can now compress earlier context with “Summarize up to here.”
• Week 19 (May 4–8): Plugin Marketplace. Plugins load from .zip archives and URLs. Anthropic has a curated catalog, claude-plugins-official, as well as a community catalog.
• Week 18 (Apr 27–May 1): claude ultrareview for CI, claude project purge, PR URL in /resume.
• Week 17 (Apr 20–24): /ultrareview public research preview, session recap, custom themes.
• Week 16 (Apr 13–17): Opus 4.7 as the default model on Max/Team Premium. The xhigh level for /effort. Interactive /effort slider. Routines on Claude Code Web. Mobile push notifications. The /usage command.
• May 6, Code w/ Claude in San Francisco: Managed Agents in research preview. Dreaming—an agent consolidates memory between sessions. Outcomes—write a rubric defining “what success means,” and a separate grader evaluates the result. Multiagent Orchestration. Webhooks for long-running outcomes. Source: the Anthropic blog.
• May 13: +50% to weekly limits through July 13, 2026 for Pro, Max, Team, and seat-based Enterprise.

OpenAI Codex (according to developers.openai.com/codex/changelog):

• May 14: Codex Mobile. Launch Codex CLI on a Mac, scan the QR code in ChatGPT mobile, and control the agent from your phone.
• May 21, v26.519: Goal Mode GA, Remote computer use, Plugin sharing.
• May 7: Codex for Chrome—a browser extension that lets the agent see what's open.
• April 9: Codex moved to token-based pricing for Business and new Enterprise plans (instead of message-based).

What this means for your choice:

• If mobile work matters to you, Codex Mobile is more mature (it pairs with a Mac via one QR code). Claude Code Web exists, but its interface is designed for a full desktop.
• If memory between sessions matters to you, Anthropic's Managed Agents with Dreaming is in research preview; Codex has nothing like it.
• If pricing for heavy autonomous processes matters to you, Codex is cheaper and Goal Mode is already generally available. Claude's /goal appeared a week earlier but is in early research access.

⚠️ One important point: Anthropic retired Output Styles in Claude Code and suggested migrating to plugins, hooks, or subagents (see the official documentation). If you're reading older articles about Output Styles, redirect that logic to the new stack of plugins / hooks / subagents.

Service availability and alternative tools#

Let's look at the options.

Codex CLI or Claude Code can be installed using the standard npm method or the official installer. Alternative #1: DeepSeek

DeepSeek V4 (released April 24, 2026) is a Chinese model; the listed payment method is Alipay (QR code).

Token prices are about 0.27 / 1.10 dollars per million. That's ten times cheaper than Sonnet 4.6 on a comparable benchmark.

Alternative #2: GLM-5.1 (Zhipu AI / z.ai)

GLM-5.1 is a 744B MoE model, MIT-licensed, capable of autonomous tasks lasting 8+ hours. z.ai has an Anthropic-compatible API endpoint, and it can technically be connected as a drop-in replacement. Pricing: a free tier through NVIDIA NIM or paid access through z.ai.

Alternative #3: Cline in VS Code

Cline is a VS Code extension and an autonomous coding agent. It supports the Cline API, ChatGPT, Claude, Gemini, and DeepSeek.

Configure it using a .clinerules file—similar to CLAUDE.md or AGENTS.md, but in Cline's format.

Alternative #4: Yandex GPT / Sber's GigaChat

Code quality is weaker than Codex or Claude, but it's enough for routine tasks (generating SQL, simple scripts). The context window is smaller, and there is no full-fledged agent harness.

Which stack to actually use:

Regardless of which tool you choose, the methodology—architecture, second brain, and verification cycle—remains the same.

Neither one is “better.” They're optimized for different scenarios. Why opinions about Codex differ so much, and how it performs in everyday work, is covered in the article on Codex reviews and experience.

How to choose for your situation: 5 typical scenarios#

Next are five archetypes, with a specific recommendation for each.

1. Complete beginner who has never written code and wants to build a first working prototype.

Recommendation: Claude Code Pro $20.

• Claude asks clarifying questions before changing code by default. That's critical for a beginner: you learn what's going on.

• Codex is autonomous by default. For a beginner, that means you'll get code you don't know how to fix if something goes wrong.

2. Marketer building a content factory (mailings, parsing, formatting).

Recommendation: Codex Plus $20.

• Most tasks are routine, with templated logic (script generation, RSS parsing, formatting, data transfer).
• Codex is cheaper for repetitive, similar tasks because it uses fewer tokens.
• Skills + Hooks let you create templates for repeatable workflows.
• AGENTS.md in the content project's repository lets you connect any AI tools in parallel.

3. Lawyer automating documents (contracts, checklist-based reviews).

Recommendation: Claude Code Pro $20 or Max 5x $100.

• Claude is better with long context and nuances (PDF parsing, summarizing contracts).
• Outcomes in Managed Agents is an ideal pattern for “check this contract against a 30-point checklist.”
• Reasoning quality matters more than price—a legal error costs more than a subscription.

4. Entrepreneur with one product (simple CRUD prototype, business logic).

Recommendation: Claude Code Pro $20 to start, then Max 5x $100 if your pace is over 30 requests a day.

• Claude asks clarifying questions, which matters for a non-programmer.
• Code quality is higher—fewer bugs at the start and easier to maintain later.
• As the project grows, Max 5x with the May boost of +50% provides a steady window without cutoffs.

5. Experienced developer or technical lead (routine tasks + complex architecture decisions).

Recommendation: both in parallel.

• Codex Pro $100 for routine work, refactoring, and code review.

• Claude Code Max 5x $100 for architecture, security-sensitive tasks, and frontend work.

• AGENTS.md as the single source of truth + @AGENTS.md import in CLAUDE.md.

⚠️ This table isn't dogma. It's a starting point. The only way to make the real choice is to test both tools in parallel on your task.

How to test both over a weekend without burning through your budget#

The fairest way to choose is to run a parallel test. The steps are described below.

1. Step 1: prepare two worktrees

   From the same original checkout, create two unused branch names from its current HEAD and place them in worktrees. Keep that checkout at the same commit while running both commands; commit the shared instructions there first if both tools must start with them. If either branch name already exists, choose an unused name instead of forcing a reset:

   git worktree add -b feature-claude ../my-project-claude HEAD
   git worktree add -b feature-codex ../my-project-codex HEAD
   

   Now you have two working copies of the same code in different directories. You can open them in two VS Code windows.

2. Step 2: create AGENTS.md and CLAUDE.md

   Before starting either agent, put the same AGENTS.md general project rules (commands, style, architecture) into both worktrees, or commit the shared rules in the original checkout before Step 1.

   In CLAUDE.md:

   @AGENTS.md
   
   ## Claude-specific
   - Sub-agent for security review
   - Permission level: workspace-write
   

   That way, the same rules work for both agents.

3. Step 3: choose one real task

   Not an example from a guide, not “calculate a factorial.” A real task you need done in your project. Something involving a database integration, authentication, or form handling. Scope it as 4–8 hours of work for an experienced developer.

   Write the task in the same format for both agents: a bullet list of requirements, acceptance criteria, and links to existing code.

4. Step 4: start them in parallel

   • Start Claude Code with the task in the first worktree.
   • Start Codex CLI with the same task in the second worktree.
   • Don't intervene for the first 30 minutes. Let them work.
   • If an agent asks clarifying questions, answer the same way in both windows.

5. Step 5: measure four metrics

   • Time to first compile / passing tests.
   • Number of iterations—how many times you had to say “not like that, redo it.”
   • Code quality—after an automatic review (npm test, lint, type-check).
   • Cost—in Codex, through /usage; in Claude Code, through /usage or the admin dashboard.

6. Step 6: repeat on 3–5 tasks

   Don't draw conclusions from one comparison. There are too many random factors. At least three different types of task: one focused on the interface, one on business logic, and one on integration with an external service.

   After 3–5 parallel tests, you'll have a real picture of “which makes your work faster.”

The test budget is about $40–80 for both tools together (if you already have Plus and Pro subscriptions, you don't need to pay extra).

Migrating CLAUDE.md → AGENTS.md without pain#

In this scenario, one canonical version plus an import is used.

Steps for a new project:

1. Create AGENTS.md in the root right away. Put everything that is universal in it:

   • Commands (build, test, lint).
   • Code style rules.
   • Architecture description (where the models are, where the services are, where the routes are).
   • Prohibitions on using certain patterns.
   • Business context.

2. Create CLAUDE.md with a thin layer of Claude-specific rules:

# CLAUDE.md

@AGENTS.md

## Claude-specific
- Run a sub-agent for security review using the `Agent` tool
- Permission level: workspace-write
- @.claude/permissions.md
- @.claude/business-context.md

3. If you have long work procedures over 500 words, move them into Skills:

   • In Claude Code: .claude/skills/<name>/SKILL.md
   • In Codex: .agents/skills/<name>/SKILL.md

4. Configure Hooks in both:

   • Claude Code: ~/.claude/hooks/ or .claude/hooks/ for the project.
   • Codex: the [hooks] section in ~/.codex/config.toml.

5. MCP servers are standard and transfer 1-to-1. If you've already configured them for Claude, copy the configuration into Codex using codex mcp add.

What doesn't transfer automatically:

• Sub-agent delegation in Claude Code has an equivalent in Codex through subagents, but the configuration is in a different format.
• @import chains are Claude-only. In Codex, the AGENTS.md hierarchy is organized through the directory tree.
• Output Style is a legacy mechanism; plugins / hooks / subagents have taken its place (see above).

If you've followed the “How to configure CLAUDE.md” guide, migration takes about half an hour: create AGENTS.md, copy the universal sections into it, and leave only Claude-specific material and @AGENTS.md at the top of CLAUDE.md.

Where the industry is headed and what that means a year from now#

If I were writing this guide a year ago, half the tools mentioned here wouldn't have existed. The same will be true a year from now. So I recommend keeping three rules in mind:

Rule 1: choose a stack you can switch in a week.

If you're on Bolt and have spent two months coding a 30-screen project there, moving away is painful. If you're on VS Code + Claude Code and want to try Codex six months from now, switching takes an hour (create AGENTS.md, import it into CLAUDE.md). That's what freedom with an open stack looks like.

Rule 2: AGENTS.md is insurance against being tied to one platform.

The open standard is read by Codex, Amp, Google's Jules, Cursor, Factory, and, since September 2026, Claude Code. The list grows every month. A file in the repository will outlast three tool changes.

Rule 3: the methodology is what shouldn't change.

Models, IDEs, prices, limits, and capabilities will change. One thing won't: how you approach a project—what architecture you use, how you give context to an agent, and what verification cycle you follow. Vibe coding like a professional = a stack built on top of a methodology. Without a methodology, the latest model produces garbage faster than usual.

What happened among competitors in May 2026, and why it changes the picture:

• OpenAI: Codex Mobile (control from iPhone), Codex for Chrome, Goal Mode GA, token-based pricing for Business.
• Google: Antigravity 2.0 on Gemini 3.5 Flash (desktop + CLI + SDK, May 19). Jules has been a full-fledged product since August 2025.
• Cursor: Composer 2.5 with Kimi K2.5 inside, ten times cheaper than Opus 4.7 (May 18).
• Anthropic: Stainless (SDK generator) acquired May 18. Plugin Marketplace becomes first-class. Managed Agents with Dreaming, Outcomes, Multiagent Orchestration.

Any one of them could change the choice in six months. Don't bet on one tool. Bet on a universal stack that can survive a change in any of them.

For further work, use Ailoot.