100 Hours Testing Claude Code vs ChatGPT Codex — Honest Results

1 Overview

▶ 0:00

Over the past years, OpenAI went from being the biggest AI company to becoming "kind of mid" in the coding space — developers basically forgot OpenAI existed thanks to tools like Claude Code. But recently, a lot of videos have been claiming that OpenAI Codex is actually better than Claude Code. Nate Herk spent over 100 hours testing both tools head-to-head to answer this question definitively.

The comparison covers three dimensions:

• Features — what each tool uniquely does best
• Price — subscription tiers, token economics, and session limits
• Three live builds — same prompts, side by side, with measured metrics

2 Claude Code Explained

▶ 0:30

Claude Code is Anthropic's coding agent. You give it a task — fix a bug, build a feature, review a pull request — and it goes off, plans the work, opens your project, edits files, runs commands, and asks for permission along the way based on your settings.

Available on:

• Terminal (CLI)
• VS Code extension
• Full desktop app (Mac and Windows)
• Web version in research preview (browser/phone)

Under the hood: Runs Opus (Anthropic's smartest model), Sonnet, or Haiku. Opus and Sonnet are top-tier for coding work.

The part Nate really likes: Claude Code is less of a tool and more of a workflow system you can shape into your own engineering setup. Key customization features:

• Skills — reusable knowledge you drop in
• Hooks — automated triggers that fire when events happen in your session
• Sub-agents — specialist agents Claude spins up on its own to handle specific work

3 Meet Codex

▶ 1:28

The new Codex (not the retired 2021 model) is OpenAI's full agentic coding system — similar shape to Claude Code but with different opinions on how work should flow.

Available on: Terminal, desktop app (Mac/Windows), VS Code extension (also works with Cursor and other IDEs), and cloud version at chatgpt.com/codex.

Models: GPT family — GPT-Codex for coding-specific work, and GPT-Codex-Spark (faster, smaller, still in research preview for Pro users).

What stands out about Codex is the unified shipping vibe. Where Claude Code feels like a workflow system you're building out, Codex feels like an opinionated machine designed to take you from "agent is done" all the way to "code is shipped to production."

• Built-in Git work trees — separate working copies so multiple tasks run in parallel without overriding each other
• Included in every ChatGPT plan — free, Plus, Pro, Business, Enterprise. Claude Code requires a paid subscription.

4 The Massive Overlap

▶ 3:24

After heavy use of both tools, the overlap is way bigger than most comparison videos admit. Both tools share:

• Edit code on your local machine
• Desktop apps and VS Code extensions
• Command line interface
• MCP support (open protocol for external tools)
• Same skills format (markdown + YAML front matter)
• Plugin marketplace for community tools
• Cloud delegation (fire off a task and walk away)
• Hooks and sub-agents

5 Claude Code's Unique Advantages

▶ 4:04

What sets Claude Code apart is the depth of customization:

• 30 hook events vs Codex's ~6 — 5× the granularity for automated triggers (on prompt submit, tool run, session start, task creation, etc.)
• Auto-delegating sub-agents — Claude spawns sub-agents on its own when needed (planner agent, explorer agent, code reviewer agent). Codex won't spawn sub-agents unless you explicitly ask.
• /ultra plan — ships the planning phase to a cloud session for browser-based review with inline comments, then sends it back to terminal for execution
• /ultra review — spins up multiple reviewer agents for deep multi-agent code review with reproduced findings (3 free runs on Pro/Max)
• /loop — recurring prompt on a schedule or maintenance mode. Handles unfinished tasks, addresses PR comments, fixes merge conflicts automatically.
• Channels — MCP server that pushes events from Telegram, Discord, or iMessage into a running session. You can text your agent from your phone.
• Claude Agent SDK — same engine as Claude Code exposed as Python and TypeScript SDK. Build your own agents on top.
• Enterprise auth — supports Bedrock, Vertex AI, and Microsoft Foundry. Codex doesn't have this level of auth flexibility.

6 Codex Fights Back

▶ 6:19

What Codex does better than Claude Code:

• Unified shipping workflow — built around work trees from the ground up. Review, stage, commit, and push from the same desktop app. A full shipping pipeline in one tool.
• In-app browser — built-in browser inside the desktop app to view your agent's work. Leave visual comments right on the page. No switching to Chrome to check results.
• Sharper computer use — particularly the QA use case: tell Codex "QA the app I just built" and it opens it, clicks around, finds bugs, logs them with severity ratings, expected vs actual behavior, reproduction steps, and triage summary.
• @Codex GitHub integration — tag @Codex in a PR comment or issue, and Codex spins up a cloud sandbox to handle it. Zero setup.
• /goal — for work too big for a single prompt but smaller than an open-ended backlog. Define a goal with a verifiable stopping condition and Codex grinds until finished (hours-long tasks). Note: Claude Code released /goal right after recording.
• GPT Image 2 access — one of the strongest image generation models, built right into Codex. Anthropic has no image generation model at all.

7 The Third-Party Harness Loophole

▶ 9:13

This is where the two companies diverge philosophically. Third-party tools like OpenClaw and Hermes Agent have blown up because they feel proactive — native crons, heartbeats, skills, and the ability to wrap coding agents.

8 Pricing Breakdown

▶ 10:22

Both tools are included with their parent subscription — no separate API key needed to start.

Context Windows

• Claude Code — 1 million tokens (Opus and Sonnet)
• Codex — ~256,000 tokens (latest GPT model)

Session Limits — The Real Issue

Many users are complaining about hitting Claude Code limits (session or weekly) faster than before. Nate tracked actual token usage in live tests and found Claude Code consistently uses 2–5× more output tokens than Codex for the same work. Since output tokens cost more than input tokens, this means:

9 Test 1: Research Report PDF

▶ 12:48

Both tools received the same prompt: build a branded research report PDF about automation tools for SMBs. No API keys were provided — both used their native web search and fetch tools.

Speed: Codex finished in ~8 minutes, Claude took 8:15. Tokens: Codex used 2.8M vs Claude's 4.7M. Codex Edge — faster and more token-efficient on the most research-heavy task.

10 Test 2: Landing Page

▶ 15:10

Same exact prompt to both tools: build a landing page with a provided logo and reference to an existing site for design inspiration.

Speed: Codex 3 minutes flat vs Claude 4:39. Visual winner: Claude Code — the underlying design and base was clearly better despite the missing logo (an easy one-prompt fix).

11 Test 3: Marketing Dashboard

▶ 16:32

The most complex test: a marketing analytics dashboard with interactive charts, date filters, and made-up but realistic data.

Claude Code Wins — roughly 4× faster, 6× fewer tokens, and visually superior. This was the most complex task and Claude's advantage was most pronounced here.

12 The Numbers — Full Breakdown

▶ 17:25

Models used: GPT-5.5 (Codex) vs Opus 4.7 (Claude Code), both on high settings.

Aggregate Results (3 builds combined)

• Total time: Codex ~26 minutes vs Claude ~15 minutes
• Total tokens: Both ~6 million total, but different distributions
• Total cost (API billing): Claude $11 vs Codex $7 — Claude was more expensive

The Output Token Pattern

The most striking finding: Claude Code's output tokens are consistently 2–5× higher than Codex across every single build:

• Test 1 (Report): Claude 84K output vs Codex 18K output
• Test 2 (Landing): Claude 80K output vs Codex 20K output
• Test 3 (Dashboard): Claude 41K output vs Codex 16K output

Efficiency Scatter Plot

When plotted on a time × token scatter (faster + leaner = bottom-left, best), Claude Code had two data points in the ideal bottom-left zone plus one outlier in the excellent zone. Codex data points clustered consistently in the middle — more predictable but less efficient.

Model Pricing Context

GPT-5.5 and Opus 4.7 have similar input pricing ($5/million), but GPT-5.5 output is $5 more expensive per million. Despite this, Codex was cheaper overall because it produces far fewer output tokens. Claude's higher output verbosity offsets its lower per-token output price.

13 The Honest Verdict — When to Use Each

▶ 21:02

This was not a clean sweep in either direction. Each tool won at different things.

14 Final Mindset — Portable Skills, Portable Folders

▶ 24:46

Because you're working with coding agents, all you're really doing is making files inside folders. Markdown, JSON, Python, TypeScript — everything gets pushed to GitHub. You can pull that same project into Claude Code, Codex, OpenClaw, Hermes Agent, or whatever the next tool is.

You're not locked in. Moving between tools is simple:

1. Open the project in the other agent
2. Tell it "I built this in Claude Code, you are Codex — walk through it, understand it, update what needs changing"
3. Swap small config files (e.g., CLAUDE.md → AGENTS.md)

Caveat: Everything discussed is as of mid-May 2026. Both tools ship at incredible speed — new models drop, pricing shifts, research preview features graduate or get removed. The architectural differences will hold up, but specific numbers might not.

🎯 Key Takeaways

⏱ Timestamp Index