Claude Code vs OpenAI Codex — Which AI Coding Agent Should You Use?

Module objectives

• Understand what a "coding agent" is in plain terms.
• See how much Claude Code and Codex actually have in common.

A coding agent is an AI that doesn't just suggest code — it does the work. You say "fix this bug" or "build me a landing page," and the agent reads your project, writes and edits files, runs commands to test its work, and checks in with you for permission based on your settings.

Both tools run almost everywhere you'd want them: a terminal version, a VS Code extension, a full desktop app for Mac and Windows, and a cloud/web mode you can drive from a browser or phone.

• Claude Code runs on Anthropic's Claude models — Opus (the most capable), Sonnet, and Haiku. Opus and Sonnet are top-tier for coding.
• Codex runs on OpenAI's GPT family, including coding-specialized models like GPT-5.5-Codex. (Note: this is not the old "Codex" model retired back in 2021 — the new Codex is a full agent system.)

Most comparison videos exaggerate the differences. In reality the overlap is large. Both tools:

• Edit code on your local machine and have desktop apps + VS Code extensions.
• Support MCP (Model Context Protocol — the open standard for connecting external tools to AI).
• Use the same Skills format: a markdown file with YAML front-matter dropped into a folder that the agent can read and invoke.
• Offer a plugin marketplace, cloud delegation (fire off a task and walk away), hooks, and subagents.

So the real question isn't "does my tool have feature X?" — it's "which one gives me the better workflow for the way I actually work?"

---

Module objectives

• Compare every plan tier and price accurately.
• Understand context windows and why the headline numbers can mislead.

Good news for beginners: both agents are included with their parent subscription — you don't need a separate, more expensive API key to start. All prices are monthly USD, last verified 2026-05-29.

| Plan | Price/mo | Coding agent | Notes |

|------|----------|--------------|-------|

| Claude Free | $0 | Claude Code not included | — |

| Claude Pro | $20 | ✅ Claude Code | Enough to explore |

| Claude Max 5x | $100 | ✅ Claude Code | 5× Pro usage |

| Claude Max 20x | $200 | ✅ Claude Code | 20× Pro usage |

| ChatGPT Free | $0 | Codex (limited trial) | A few sessions/week |

| ChatGPT Go | $8 | Codex | Limited |

| ChatGPT Plus | $20 | ✅ Codex | Baseline |

| ChatGPT Pro ($100) | $100 | ✅ Codex | 5× Plus limits |

| ChatGPT Pro ($200) | $200 | ✅ Codex | 20× Plus, near-unlimited feel |

Footnote sources (these resolve to live vendor pages):

• Claude plans & pricing — claude.com/pricing [^1]
• ChatGPT plans & pricing — chatgpt.com/pricing [^2]
• Codex pricing — chatgpt.com/codex/pricing [^3]

Two correction notes versus older comparisons: the $100 ChatGPT Pro tier is a permanent plan (added April 2026), not a temporary promo as some 2026 videos described — though a launch promotion did temporarily double Codex usage on that tier through May 31, 2026. If you plan to test Codex heavily, that $100 tier is currently one of the best values in the AI-coding market.

One more economic wrinkle (verified): OpenAI permits routing your ChatGPT subscription through third-party agent wrappers like OpenClaw or Hermes — OpenAI's CEO publicly endorsed this on May 2. Anthropic's Agent SDK terms disallow reselling Claude.ai login or rate limits to third-party products unless pre-approved. If you live inside those third-party tools, that favors Codex.

A context window is how much text — your code, instructions, and the conversation — the model can "hold in mind" at once. Bigger generally means it can reason over more of your project before losing track.

| Tool | In-product context | Notes |

|------|--------------------|-------|

| Claude Code (Opus 4.6+/Sonnet 4.6) | 1,000,000 tokens | Generally available since March 2026, no surcharge |

| Codex (GPT-5.5-Codex) | 400,000 tokens | The Codex API supports ~1M, but the Codex app is capped at 400K |

Heads-up: some 2026 videos quote Codex at "256K" — that was the previous GPT-5.4-era figure and is now out of date. The current verified number inside Codex is 400K.[^4] Claude Code holds roughly 2.5× more context in-product. (Separately, some users feel they get more work done in Codex before hitting usage limits — that's an anecdotal token-efficiency impression, not a context-window fact, so don't conflate the two.)

---

Module objectives

• Learn where each tool genuinely shines.
• Use a decision flowchart to pick the right tool for a task.

The tools have different shapes:

• Claude Code = a customizable workflow system. Its standout strength is depth of customization. It exposes 30 distinct hook events (verified against the official docs) — automated triggers that fire at moments like a session starting, a prompt being submitted, or a tool finishing. Codex exposes far fewer — around six by one count from the source video, though that figure isn't confirmed in OpenAI's docs, so treat the gap as directional (Claude Code clearly offers more hook granularity) rather than an exact ratio. Claude Code can also auto-delegate subagents (spin up a planner, an explorer, a reviewer on its own), supports enterprise hosting (Bedrock, Vertex AI, Microsoft Foundry), and ships the Claude Agent SDK for building your own agents.

• Codex = an opinionated shipping machine. It's built around git worktrees (separate working copies so multiple tasks run in parallel without colliding), with review, commit, and push in one desktop app. It has a polished in-app browser for QA-ing what you just built, sharp computer-use testing, a smooth @Codex GitHub-mention workflow, and built-in image generation (Anthropic has no image model of its own).

Both recently shipped a /goal command for long-running, autonomous objectives — Claude Code added it in version 2.1.139 on May 12, 2026.

On "feel": the source creator says Claude Code feels more creative and pushes back on bad ideas, while Codex feels better at following instructions and reviewing code. He's explicit that these are gut impressions, not measured benchmarks — treat them as opinions, not data.

A practical, increasingly popular pattern is plan in Claude Code, execute in Codex: brainstorm and shape the approach with Claude Code's creativity and push-back, then hand the plan to Codex to ship. Because your project is just files in folders pushed to GitHub, switching tools is cheap (mostly swapping a CLAUDE.md for an AGENTS.md) — you're never locked in.

flowchart TD
    A[What's the task in front of you?] --> B{Planning or executing?}
    B -->|Planning, brainstorming, a hard problem| C[Lean toward Claude Code]
    B -->|Executing a known plan, shipping| D[Lean toward Codex]
    C --> E{Need 30 hook events, auto-delegating subagents,
or enterprise Bedrock/Vertex/Foundry auth?}
    E -->|Yes| F[Claude Code is the clear pick]
    E -->|No| G((Either works well — try the
plan-in-Claude, execute-in-Codex combo))
    D --> H{Need worktree shipping, in-app browser, @Codex PR review,
built-in image generation, or OpenClaw/Hermes subscription routing?}
    H -->|Yes| I[Codex is the clear pick]
    H -->|No| G
    F --> G
    I --> G

The honest bottom line: it's rarely "which tool is better" in the abstract — it's "which tool is better for the task in front of you." If you already pay for one, you already have a top-tier coding agent. If you can spare it, subscribing to both ($20 + $20) and using each for what it does best is a legitimately strong setup.

[^1]: Claude consumer pricing — https://claude.com/pricing (verified 2026-05-29).

[^2]: ChatGPT pricing — https://chatgpt.com/pricing (verified 2026-05-29). Covers every ChatGPT tier in the table above, including Free, Go ($8/mo), Plus, and both Pro tiers.

[^3]: Codex pricing — https://chatgpt.com/codex/pricing (verified 2026-05-29).

[^4]: Codex GPT-5.5 400K context, official OpenAI Codex repository — https://github.com/openai/codex/issues/19464 (verified 2026-05-29).