Claude Code vs Codex: Claude Code scores 62 on Artificial Analysis, and Codex gets within 8 points for 87% less per task

Is Codex better than Claude Code?

No. Claude Code beats Codex by 7 points on every coding agent test Artificial Analysis runs, and Codex wins on cost and speed. Artificial Analysis runs both agents on the same tasks, with Anthropic's and OpenAI's top models at maximum effort. We read its Claude Code vs Codex page on 1 November 2026:

The index is the average of the three tests below it, each rounded to a whole number on that page. FrontierCode asks if a maintainer would merge the agent's change to a real repo. Terminal-Bench gives it jobs to do in a shell, and SWE-Atlas asks questions about how a codebase works. Claude Code wins all three. It also costs more. Here's every model Artificial Analysis lists for each agent from Anthropic and OpenAI:

GPT-6.1 Sol is the row to look at. It scores one point under Astra for about a fifth of Astra's cost, and every Claude Code run that beats it costs at least $11.42. So Claude Code's 8 extra points cost 7.6 times as much per task, which makes Codex 87% cheaper. Claude Code also works longer. Its Opus 5.5 run took an hour and 146 turns per task, against 25 minutes and 39 turns for Codex on Astra, and it used 4.5 times the tokens.

Sonnet 5.5 is the trap. Anthropic charges half as much per token for Sonnet as for Opus, yet in Claude Code it cost more per task, $13.30 against $11.42, because it burned 24.1 million tokens to finish. Artificial Analysis prices every task at pay-per-token API rates. So the cost column is what an API key pays, not what a plan costs. If you're on a subscription, tokens per task tells you more about how fast you'll reach your limit.

How much do Claude Code and Codex cost each month?

Both start at $20 a month: Claude Code on Claude Pro and Codex on ChatGPT Plus. Only Codex has a free tier, and it's a small one. We read every row below from Anthropic's pricing page and OpenAI's Codex pricing page on 1 November 2026. US list prices, before tax:

At $100, only OpenAI drops the 5-hour window. At $200, both do. Anthropic renamed Max 20x to Ultra on 30 October 2026, and its Ultra help article says: "The Max 20x plan is now called the Ultra plan." The same article adds, "Verified Ultra accounts don't have a five-hour limit." OpenAI's page says the same of every Pro tier: "Pro plans currently have no five-hour limit." Max, at $100, keeps its window.

The old name said 20x. Anthropic now sizes Ultra at 10 times Pro's weekly usage, at the same $200. OpenAI doesn't print a multiple for any Pro tier on its Codex page. Neither vendor publishes a weekly cap in messages. The full plan math for each side is in Claude Code pricing and ChatGPT Plus vs Pro, and the two $100 to $200 tiers meet head to head in Claude Max vs ChatGPT Pro.

Which has better usage limits, Claude Code or Codex?

Codex has the limits you can plan around, because OpenAI prints a range per model and Anthropic prints none. For Plus and Standard Business, OpenAI estimates 15 to 160 local GPT-6.1 Sol messages per 5 hours, 5 to 45 on GPT-6 Astra and 350 to 3,000 on GPT-6 Luna. It also warns, "These estimates are not fixed message limits." Anthropic sizes each Claude plan as a multiple of Pro and never says what Pro holds.

OpenAI meters Codex in credits per million tokens. GPT-6 Astra costs 250 credits per million input tokens and GPT-6.1 Sol costs 50, so Astra drains a plan five times faster per token. Cloud tasks use more of the allowance than local ones, per the same page. Codex shares its pool with ChatGPT Work. OpenAI doesn't list ordinary ChatGPT chat as part of it.

Claude Code shares its pool with Claude chat. Anthropic's Pro and Max help page says the limits "are shared across Claude and Claude Code, meaning all activity in both tools counts against the same limits." A long chat in the morning leaves less for the terminal. That same page says the monthly API credits on Ultra "can't be used in Claude Code." How each window resets, and what each wall message means, is in Codex usage limit and Claude Code usage limit.

How do Claude Code and Codex work differently?

Claude Code and Codex run in the same places: a terminal, an IDE, a desktop app and the web. They differ in defaults. We compiled this from Anthropic's Claude Code docs, OpenAI's Codex docs and the openai/codex repository, read on 2 November 2026:

The sandbox row matters most if you let an agent run unattended. Anthropic's sandboxing page says "The sandbox is off by default," so Claude Code relies on its permission prompts until you run /sandbox. On native Windows it can't sandbox at all: it "runs commands unsandboxed." OpenAI's sandboxing page says "The default permissions mode applies sandboxing automatically." On macOS that needs no setup, Linux needs bubblewrap installed first, and Windows gets its own sandbox in PowerShell.

The instruction row makes switching cheap. Claude Code's docs say it can read an AGENTS.md "in place of a CLAUDE.md," so one file in your repo briefs both agents. Codex is also the one you can read. Its CLI is open source under Apache-2.0, with 132.8k stars on GitHub. Claude Code's strength is the session itself. It plans, edits across files and checks its work over many more turns, which is where the 146 against 39 above comes from.

Both also take work you hand off and check later. Claude Code on the web lets you "kick off long-running tasks and check back when they're done" at claude.ai/code. Codex Web at chatgpt.com/codex is OpenAI's cloud agent, and OpenAI's pricing page warns that its cloud tasks "may use more of your allowance than local messages."

Should you pick Claude Code or Codex?

Pick Claude Code if you want the highest score and you're paying a flat plan. Pick Codex if you pay per token or want the sandbox on from the start. Here's how that plays out by budget.

1. $0. Codex. ChatGPT Free gets GPT-6 Luna in the desktop app where it has rolled out, and Claude Free has no Claude Code.
2. $20. Either. Claude Pro gives you Opus 5.5, the top scorer, with a 5-hour wall shared with chat. ChatGPT Plus gives you GPT-6.1 Sol, 8 points lower, with a published range per window.
3. $100. Codex, if the 5-hour wall is what stops you. ChatGPT Pro $100 has none, and Claude Max still does.
4. $200. Either. Both drop the 5-hour window now (Ultra once Anthropic verifies your account), so pick on the agent you prefer.
5. An API key. Codex on GPT-6.1 Sol, at $1.50 a task against $11.42 for Claude Code on Opus 5.5.

Claude Code loses on cost per task. Codex loses on quality: its best score, 55, sits 7 points under Claude Code's 62. You can run both, since one AGENTS.md serves the pair, and a $20 plan of each costs $40.

Whizi, which publishes this page, includes neither Claude Code nor Codex. It's a chat app, so it won't run in your terminal or edit your repo. What Whizi Pro does give you is Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol in one chat for $29.99 a month, from 2,500 credits. That's more than Claude Pro or ChatGPT Plus alone, so if you want an agent, buy one of those instead.