Claude Code vs Cursor vs Codex in Practice
Sixteen subagents died at once on a 429. We compare the three tools on instruction files, hooks and review output, not feature lists.
PR: This article contains placements for affiliate programs we plan to join. Rankings and verdicts do not depend on payouts. We found no affiliate program for Claude Code, Cursor or Codex.
Feature tables are everywhere, so this comparison looks at operations instead: where the instruction file lives, whether rules can be enforced, and how things break when you run agents in parallel. Specs and prices were checked on the official pages on October 16, 2026.
Types and the incidents every tool shares are covered in Choosing a Coding Agent.
Versions we installed
On our test box (Linux, Node.js v20.19.2), using the official npm packages, checked 2026-10-16. Watch the package names: the bare npm packages claude (self-described as “not the official Claude Code package”) and codex (an unrelated docs generator) are different software, so npx claude runs the wrong thing.How to install the right package is in Getting Started with Claude Code.
$ npx @anthropic-ai/claude-code --version
2.1.301 (Claude Code)
$ npx @openai/codex --version
codex-cli 0.163.0We did not install the Cursor CLI this time. Cursor details come from its official pages only.
Instruction files
| Tool | File | Source (checked 2026-10-16) |
|---|---|---|
| Claude Code | CLAUDE.md; reads AGENTS.md when no CLAUDE.md exists (v2.1.277+) | https://code.claude.com/docs/en/memory |
| Codex | AGENTS.md, nested per directory to limit injected context | https://developers.openai.com/codex/pricing |
| Cursor | .cursor/rules/ (listed as an import source by Claude Code’s /init) | https://code.claude.com/docs/en/memory |
If you use more than one tool, keep shared rules in AGENTS.md and import it from CLAUDE.md with @AGENTS.md, as the Claude Code docs suggest.
How far we cut ours: trimming CLAUDE.md.
Enforcement: hooks
Claude Code’s docs say CLAUDE.md is context, not enforced configuration; to block an action, use a PreToolUse hook (https://code.claude.com/docs/en/memory, checked 2026-10-16).
We check commit format with a hook. The hook itself misfired: heredoc commit messages reached it truncated at 200 bytes. We now write the message to a file and pass it with git commit -F.
printf '%s\n' "fix: validate input" "" "details..." > /tmp/msg.txt
git commit -F /tmp/msg.txtExit codes and the commands that slipped through: Claude Code hooks in practice.
Parallel runs: the API sets the ceiling
We started 16 subagents and all of them died on a 429 at nearly the same moment. Restarting all 16 repeats the failure. Now we resume only the missing ones and launch in two waves of 8–10.
Claude Code’s docs list recommended TPM/RPM per user by team size (1–5 users: 200k–300k TPM, 5–7 RPM) (https://code.claude.com/docs/en/costs, checked 2026-10-16). Codex publishes estimated messages per five hours for Plus and Standard Business; Pro currently has no five-hour limit (https://developers.openai.com/codex/pricing, checked 2026-10-16).
How we recover the ones that died: subagent failures.
Review: fix the output format
When a second model reviews code, make it label each earlier finding as keep, downgrade or withdraw. Without that, old findings come back every round dressed as new ones. Grep any facts in the brief before you send it. Codex ships a review subcommand (seen in codex --help, v0.163.0).
How many rounds to run: when to stop review.
For using a second model to rebut a design before any code exists, see Have Another AI Rebut Your Design.
Entry prices (as of 2026-10-16)
| Tool | Individual plans | Source |
|---|---|---|
| Claude Code | Pro $20/mo ($17/mo billed annually), Max from $100/mo | https://claude.com/pricing |
| Codex | Plus $20/mo; Pro $100, $200 or $500/mo | https://developers.openai.com/codex/pricing |
| Cursor | Plus $8, Premium $30, Super $100, Ultra $200/mo (sold as the X Subscription) | https://cursor.com/pricing, https://cursor.com/docs/models-and-pricing |
As of October 2026, Cursor is sold as part of X’s “X Subscription”, and the pricing page also lists Grok and X features. Two caveats: Plus ($8) includes Grok models only, and on Premium and above, third-party models (Claude, GPT and others) are not included in the flat price; they are billed at API list price from prepaid credit. The former Pro ($20) and Pro+ ($60) plans continue only as legacy plans for existing customers (https://cursor.com/help/account-and-billing/pricing, checked 2026-10-16). If you mainly want Claude models, don’t compare on the headline price alone.
All prices exclude tax. Prices change every few months; check before you subscribe. More on Claude Code costs in What Claude Code Really Costs.
Summary
- Keep shared instructions in one AGENTS.md.
- Put must-follow rules in hooks, not instruction files.
- Size parallelism from your API limits; resume only what failed.
- Have reviewers answer in keep / downgrade / withdraw form.