Claude Code vs Codex: Stop Picking a Side, Start Picking a Task

A developer-authored comparison argues that Claude Code and OpenAI's Codex are effectively tied at the model layer, with the real differences lying in pricing, token consumption, and the programmability of the surrounding harness. Citing a controlled Composio head-to-head on Opus 4.7 versus GPT-5.5 and vendor-published benchmarks, it reports that Claude Code spends more tokens on tool-heavy MCP work while offering deeper customization through subagents, Skills, hooks, and dynamic workflows, whereas Codex spans six surfaces under one product.

In January 2026, Andrej Karpathy posted a thread about going from 80 percent manual coding to 80 percent agent coding in a single month. It pulled 40,000 likes in a week, and the replies split exactly down the middle: half pointing at Claude Code, half pointing at Codex. Six months later, that split is still the live debate, and most of the content about it is trash. It is either written by someone affiliate-linked to one answer, or it repeats folklore numbers that controlled tests have already debunked. This article is the opposite of that. I have not run a paid 100-hour head-to-head myself, and I will not pretend I did. Every number below comes from public benchmarks, vendor announcements, and long-form comparisons by developers who did run the tests, all linked inline. The short version: the model layer is a tie, the pricing is not, and the right answer depends on the kind of work you are doing, not on which company you like. Start with the number that decides most purchases before any benchmark gets read. Both tools sit on top of a chat subscription. Codex comes with ChatGPT plans, Claude Code comes with Claude plans, and the tiers look similar on the surface: roughly $20, then $100, then $200. Under the surface they are not similar at all. So the honest statement is: Codex is cheaper to try, and both get expensive at scale. If you are on a $20 budget, this section is probably the whole decision for you. If money is not the constraint, read on, because the free-tier math is not where the interesting differences live. Plan pricing is the visible number. The number that decides whether you stay inside your limits is tokens per task, and Claude Code consistently spends more of them. It reads more files, plans before writing, and verifies tools before calling them. That behavior has a price. The cleanest controlled test I found is a head-to-head from Composio https://composio.dev/content/claude-code-vs-openai-codex that ran the same two prompts a PR-triage system and a real-time code review UI against Claude Code on Opus 4.7 and Codex on GPT-5.5, same machine, same MCP setup. The results kill two myths at once: There is a pattern underneath: the gap widens when the agent is doing tool-heavy work. If your session talks to Linear, GitHub, and a database through MCP, Claude Code's "check tools first, then plan" loop runs the bill up faster. For a self-contained refactor with no tool calls, the gap nearly closes. If you use Claude Code and want to cut consumption, Firecrawl's token efficiency guide https://www.firecrawl.dev/blog/claude-code-token-efficiency documents 12 techniques that benchmarks show cutting costs by 77 to 91 percent. On public benchmarks, the two model families are within noise of each other. From the vendors' own published numbers: Arguing about which model is 0.4 benchmark points ahead is a waste of your attention. All these numbers are vendor-reported, they move every quarter, and the interesting variable is not the model. It is the harness around it. This is where the two products genuinely diverge, and it is the part most comparisons gloss over. Claude Code is the deepest programmable harness on the market. The stack includes Subagents https://code.claude.com/docs/en/sub-agents specialized agents in .claude/agents/ with their own context window and tool allowlist , Skills https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills the SKILL.md format Anthropic open-sourced in December 2025 , Hooks https://code.claude.com/docs/en/hooks 26 lifecycle events you can intercept with shell scripts , a plugin marketplace, and Dynamic Workflows, where Claude orchestrates tens of subagents in one session. If you want a custom policy like "run the linter before every commit and reject commits that fail tests without human approval," hooks give you exactly that knob. Codex takes the opposite bet: one product across six surfaces. CLI, IDE extension, Codex Cloud https://chatgpt.com/codex , the ChatGPT app sidebar, mobile GA May 2026 , and a Chrome extension. All of them share your ChatGPT account, your session history, and your AGENTS.md config. You can start a refactor on your phone during a commute, pick it up in VS Code at your desk, and review the PR from the Chrome extension without losing state. Codex is now used by "more than 5 million people every week" https://openai.com/index/openai-frontier-models-and-codex-are-now-available-on-aws/ per OpenAI's June 2026 announcement. It is also open source under Apache-2.0 https://github.com/openai/codex , so you can read and modify the harness itself. The sandboxing philosophies are the deepest split. Codex enforces isolation at the kernel layer: Seatbelt on macOS, bubblewrap with Landlock on Linux, the Windows sandbox in PowerShell, with network off by default. The OS enforces the boundary before the model ever reaches it, which is deterministic. Claude Code enforces policy at the application layer through those hooks plus an Auto mode classifier https://www.anthropic.com/engineering/claude-code-auto-mode shipped in March 2026, which reviews tool calls. Less deterministic, far more expressive. If you are running an agent on a sensitive codebase and want a hard guarantee it cannot touch the network, Codex's model is what you want. If you want to encode your team's policy as code, Claude Code's is. Codex uses AGENTS.md https://agents.md/ , the community-defined repo instruction file. Cursor, Windsurf, OpenCode, and most other agents respect it. One file, many agents. Claude Code uses CLAUDE.md, which is proprietary to Anthropic's ecosystem but more powerful inside it: hierarchical resolution where the most specific file wins, @path imports so you can compose instruction files, and auto-memory that writes back when you tell Claude to remember something. The practical catch, noted in Builder.io's comparison https://www.builder.io/blog/codex-vs-claude-code : Claude Code still does not read AGENTS.md, so multi-agent repos end up maintaining two files. Here is the scannable version, so you can save this and skip the rest of the internet's take: If I were setting up from scratch today, I would not pick one. The people running both consistently report the same hybrid pattern: Claude Code for interactive work in the terminal, where the fast feedback loop and the hook system earn their keep, and Codex Cloud for background tasks, long autonomous runs, and PR review through @codex mentions. Both speak MCP, so one MCP server setup works in either. The cost of running both at entry level is one $20 subscription, which is less than the hourly rate of the debugging session you will need after picking a tool by brand loyalty instead of by task. The debate is loud because the models are tied and the differences are architectural. Architectural differences are good news. It means the answer is not "which one is better," it is "which one matches the work in front of you this week." I write about AI tooling, backend engineering, and the developer workflow every week. Subscribe, it is free. Which one is in your terminal right now, and what made you pick it? Did the $20 pricing asymmetry factor in, or did you choose on harness features? Tell me in the comments.