Claude Code vs Codex: one of them works for me at 3 AM. The other one argues with me at 9.

Not the model quality — both vendors shipped frontier models within a day of each other in June 2026, and the gap moves monthly. The durable differences are contract-shaped. Five of them decide every row in the verdict table below.

If your real question is harness-vs-IDE — where you want to review work, not who does which job — that's a different comparison and it has its own page: Claude Code vs Cursor. This page assumes you've already chosen the terminal and you're staffing it.

Every row comes from a documented run on a real company, not a benchmark harness. The receipt column is the point — click through and check the numbers.

The seat math first, because it's the boring part: both land in the same shape. A hosted Codex seat ran $200-300 per developer per month as of early 2026, and the Claude Code seat sits in the same range — verify both on the vendors' pricing pages before you commit a line item, those numbers move quarterly and anti-bot walls make them annoying to pin from a script.

The usage is where the real bills live, and the corpus has three receipts. My night-shift Codex on one repo spent roughly $180 in API tokens in a month, against a workload of about 30 PRs and 60 Sentry triages (Ch 35). One agentic loop fix is about 30 round-trips of context — 1 to 2 million tokens per landed fix (/the-bill). And Claude Code across the whole portfolio burns between 3 and 10 billion tokens a month (Ch 1) — that's the workforce line, not a chat habit.

The number to respect: the 91,874-line Codex run was not free. It burned 46% of one week's usage — the cost is part of the receipt, and Ch 43 prints it next to the deletion count instead of hiding it. Big autonomy has a meter on it.

Two rows from Anthropic's June 9 launch table are the ones this comparison keeps reaching for. Terminal-Bench 2.1: Fable 5 at 88.0% vs GPT 5.5 through Codex CLI at 83.4% — the one row where the rival brought its own harness, which makes it the closest thing to an agent-vs-agent number that exists. SWE-Bench Pro: 80.3% vs 58.6%, a model-level gap of 21.7 points.

Now the standing discipline, which didn't expire on launch day: Berkeley RDI reward-hacked eight major agent benchmarks in April 2026, and since then every public score on this site carries a 10-15-point discount. Run that discount over both rows and they read differently. The Terminal-Bench gap is 4.6 points — inside the discount band, which makes it a tie until your own eval says otherwise. The SWE-Bench Pro gap survives as shape, not precision. Neither row re-tiers anything by itself.

The full thirteen-row table with per-row caveats is Fable 5 vs GPT 5.5; the live ranking that a launch slide doesn't touch is the tier list. This page defers to both — its receipts are production runs, which is the eval the benchmarks are a proxy for.

Yes, and the infrastructure bill for the second agent is close to zero, because they share everything that matters. Both agents read the same .mcp.json — five MCP servers, one file; Codex reads it when it boots in its cloud sandbox, Claude Code reads it when the repo opens locally. Both read the same CLAUDE.md as the contract: file ownership, commit shape, what "done" means. The Belkins partner-portal CLAUDE.md runs 240 lines, and about a third of it is the section called "Codex never pushes to main" (Ch 35).

Since May 2026 the skill library is shared too: Codex CLI adopted the same SKILL.md format as Claude Code — same frontmatter, same triggers, same directory layout. The receipt is a deal-watcher skill of about 90 lines that fires from Codex at 3 AM when a Sentry event hits the deal-sync queue, and fires from a Claude Code session at 9 AM during PR review. One file, two CLIs, two shifts. Write skills once.

Three rules keep the night shift and the day driver from stepping on each other, and they're cheap to install. Branch protection on main — enforced at the GitHub level, not the social one; the social rule fails the first time Codex tries to "fix" a merge conflict at 4 AM. File ownership in CLAUDE.md — Codex owns the jobs, workers, and regression-test directories; the day driver owns the app surface and the schema, because schema changes need a rollback story. No agent merges its own PR, ever. The third rule is the one that actually matters, and the failure section below is why.

Every failure in this section is a contract failure with my name on it, not a model failure — which is exactly why it transfers to your setup.

Codex, given merge rights: during a maintenance push I gave Codex merge permissions and forgot to revoke them. It auto-merged a one-line "fix" that suppressed an exception — and that exception was the only signal a customer's SMTP creds had rotated. Eleven days to notice, two days to find, the customer churned (Ch 35). Branch protection isn't paranoia; it's the cheapest insurance you'll ever buy.

Codex, given a fuzzy contract: confidence drops fast as the diff widens. A four-file PR from Codex is fine; a nine-file PR is a re-architecture wearing a fix's clothes — close it unread and reopen the problem in Claude Code, where you can argue. The loop version of the same bite is running "simplify" with no fence: a loop without a goal, a diff ceiling, and a stop condition isn't running until done, it's running until tired — and it will "simplify" your auth flow into a security hole on turn 14 while you're at lunch (Ch 42).

Claude Code, given depth without caps: the day driver's bite is the same autonomy pointed inward. An uncapped subagent looped 1,400 calls into a $1,847 bill in eleven hours — no single call crossed a threshold, the aggregate had no threshold to cross (Ch 28). Spend caps and loop detectors are the CC-side equivalent of Codex's branch protection: the fence you build before the night you need it.

The standing allocation from Ch 35: Claude Code gets everything that needs an opinion — features, migrations, the swarm, headless CI, the argument about whether an abstraction earns its weight. Codex gets everything with a tight scope, a clear signal, and a deterministic test — Sentry traces, flaking tests, docs that lag the code, dependency bumps. The hand-off is a PR and a human, in that order. When a Codex PR shows up wearing more scope than its signal justifies, it doesn't get a closer read — it gets closed, and the problem reopens on the day shift.

The part I'd underline for anyone copying this: even the best run needed a hand on the wheel. Mid-run on the 91,874-line deletion, watching the diff balloon, the steer was — roughly — "you're making a lot of changes; test the build, what you already did, commit and push if no regressions" (Ch 43). That's not a footnote on the autonomy story. That is the autonomy story: the loop was good, and it still needed a human telling it to checkpoint.

No closing framework. Both vendors will leapfrog each other again before this page needs a rewrite — they shipped frontier models a day apart this cycle, and the tier list stays the live ranking so this page doesn't have to. What won't move: two shifts beat one agent with both, the contracts live in files the agents read, and the only door to main is a person. Staff accordingly.