Codex vs Claude Code, We Run Both Daily

Codex vs Claude Code, in October 2026: Codex (OpenAI, GPT-6.1 Sol) is the cheaper, steadier agent for work you can describe in a paragraph. Claude Code (Anthropic, Opus 5.5 by default) is better when a change spreads across many files. Both now ship hooks and subagents. We run both, and route between them by rule.

Updated October 19, 2026 · first published August 5, 2026 · by the AITerm team

What are Codex and Claude Code, exactly?

Both are terminal agents. You run a command inside a project folder, describe the task in English, and the agent reads files, runs your build, proposes a diff and waits for approval before writing. Same loop, two houses.

Codex CLI is OpenAI's first-party agent, written in Rust and shipped as a single fast binary. It reads AGENTS.md for project conventions and applies edits through an apply_patch primitive rather than free-form file writes. Claude Code is Anthropic's, and it grew the wider extension surface first: skills that load on demand, subagents with their own context windows, hooks, plugins.

The short version of the verdict fits in two lines. Codex is the cheaper, steadier agent for work you can scope in one paragraph. Claude Code is the one you want when a change touches nine files and you cannot fully describe it up front.

# both, side by side, on the same Mac
npm install -g @openai/codex        # or brew install codex
curl -fsSL https://claude.ai/install.sh | bash

cd ~/code/my-project
codex                                # OpenAI agent
claude                               # Anthropic agent

You're probably thinking this is a sales pitch for our terminal. Fair. Everything in the comparison sections works in any terminal, including the one you have open right now. We say where our product enters the picture, and it is near the end.

One correction since we first published this in August. We used to call Gemini CLI the third agent worth installing. Google stopped serving it to personal Pro, Ultra and free accounts on June 18, 2026 and pointed people at the Antigravity CLI instead, which we cover in Claude Code vs Gemini CLI and Antigravity vs Claude Code. For most individual developers, this is now a two-horse race.

Codex vs Claude Code, six things that changed since August

Most pages ranking for this comparison were written against models that no longer run by default. Here is what moved between our first version and this one, all checked against the vendors' own docs in October 2026.

Look, the honest reading of that list is convergence. Every structural feature we used to cite as a reason to pick Claude Code now has a Codex equivalent with a different file format. What still separates them is which model is behind the prompt, how each one bills you, and how each one stops a bad command.

The benchmark numbers are a generation behind

Every score you will see quoted in this debate was measured on a model the agents no longer run by default. Our August version cited GPT-5.3-Codex at 77.3% on Terminal-Bench 2.0 against Claude Opus 5 at 89.1% on Terminal-Bench 2.1. Two versions of the benchmark, two model generations ago.

The October picture is not cleaner. On the Terminal-Bench 2.1 board mirrored by BenchLM on October 17, 2026, OpenAI's best entry is GPT-5.6 Sol at 91.9%, Anthropic's best Opus is Opus 4.8 at 74.6%, and there is no entry at all for Opus 5.5 or GPT-6.1 Sol. Comparing those two numbers tells you about older models, not about the agent you would install today.

The measurement worth taking seriously is cost per unit of work, because you can reproduce it. On one matched task reported by Firecrawl in June 2026 and relayed by Tech Insider, Claude Code used about 192,000 tokens and $2.50, against 136,000 tokens and $2.04 for Codex. A 1.4x token gap and a 23% cost gap, on the previous model generation.

That gap has a cause, and model quality is the smaller part of it. Claude Code's default loop checks its tools, plans, then writes. Codex scopes tighter and goes straight at the file. On tool-heavy work with several MCP servers connected, the first loop costs real money. On a contained refactor, the difference nearly vanishes.

How much does each one actually cost?

Codex is cheaper to start and Claude Code is cheaper to predict. Codex rides inside ChatGPT plans, from a free tier on a lighter model up to a $500 Pro tier. Claude Code sells through Claude subscriptions, from Pro at $20 a month, and the free Claude plan does not include it.

Prices from OpenAI's Codex pricing page and claude.com/pricing, checked October 19, 2026. The per-model message ranges are where Codex gets interesting, and we broke them down in Codex usage limits explained and Codex pricing.

The number that reframed this for us is in Anthropic's cost documentation: across enterprise deployments Claude Code averages about $13 per developer per active day, or $150 to $250 a month, with 90% of users under $30 per active day. Read that next to a $200 subscription and the pricing debate stops being about the sticker.

Honestly, we spent a month optimising the wrong variable here. We chased the cheaper tier while the actual cost was a senior engineer reviewing agent output at 6pm. Turns out the plan is the small line in that budget.

Where each agent runs out of context

Context is what actually ends a session. Anthropic's model configuration docs say Opus 5.5 and Sonnet 5.5 run with a 1 million token window by default, no [1m] suffix needed, and auto-compact at about 967,000 tokens. OpenAI's models page does not publish a window size for GPT-6.1 Sol, so we stopped quoting one.

Bigger is not automatically better. A compacted session is a summarised session, and summaries lose the detail that made the agent right about your codebase forty turns ago. In practice we clear far more often than either tool forces us to.

Subagents are the other half of this, and both tools now have them. Claude Code documents them as separate context windows that do the noisy work, a codebase search or a log trawl, and hand back a summary. Codex's built-in explorer plays the same role, and a custom TOML agent can pin a cheaper model so the search does not burn your main window.

# Claude Code, before a long task
/context          # what is filling the window right now
/clear            # cheaper and more accurate than compacting
/model opusplan   # Opus to plan, Sonnet to execute

# Codex CLI, same reflex
/compact          # explicit, before it happens to you
/new              # fresh session, AGENTS.md reloaded

# .codex/agents/searcher.toml, a cheap read-only scout
name = "searcher"
description = "Finds where things live. Never edits."
developer_instructions = "Search the repo and report file paths with one line each."
model = "gpt-6-luna"

So, the honest framing: Claude Code has more rope by default. Codex makes you think about it sooner. Both punish the same mistake, which is running a fourth unrelated task in a session opened three hours ago.

AGENTS.md and CLAUDE.md are not the same file

Each agent reads its own project memory file at session start, and neither reads the other's. Codex loads AGENTS.md, Claude Code loads CLAUDE.md. The content overlaps by maybe 80%, which is exactly what makes this trap work.

Our first attempt was two full files, maintained in parallel, both authoritative. Within three weeks they had drifted: AGENTS.md still described a build script we had renamed, and Codex kept calling it. Nobody noticed, because a stale memory file never fails loudly. It just makes the agent quietly wrong.

The fix took four lines. One canonical file, one pointer.

# AGENTS.md (repo root)

The project rules live in CLAUDE.md at the repo root.
Read that file first and follow it exactly.

Codex-specific notes:
- Sandbox: workspace-write is enough, this repo never needs network
- Run ./build.sh before proposing a diff, it is fast

The rules themselves are short and every line changes behaviour. On the AITerm repo that means the build command, one hard prohibition (never push without an explicit ask), and the localisation rule that every visible string goes through L("key") in all seven languages. Both agents break that last one on their first day if you leave it out, and both respect it once it is written down.

Codex memories add a wrinkle. They are off by default and live in ~/.codex/memories/ as generated state, which OpenAI says not to edit by hand. We leave them off on shared repos: a rule that matters belongs in the file your teammates can read in review.

Which one is safer to leave unattended?

Codex, by a narrower margin than in the spring. It asks you to pick two independent things: a sandbox mode (read-only, workspace-write or danger-full-access) and an approval policy (on-request, never or granular). Filesystem reach and interruption frequency stop being the same dial, and workspace-write keeps the network off unless you turn it on.

In a version-controlled folder Codex starts on its Auto preset, which is exactly that pair: workspace-write plus on-request. You can also route escalations to a reviewer agent with approvals_reviewer = "auto_review", at the price of extra model calls on your plan.

Claude Code gets to a similar place through permission modes, settings and hooks. Codex now has hooks as well, and its docs say something we wish more people repeated: treat tool hooks as a useful guardrail, not a complete enforcement boundary. A PostToolUse hook cannot undo a command that already ran, on either side.

# Codex, write to the workspace, network stays off
codex --sandbox workspace-write --ask-for-approval on-request

# Codex, block a command before it runs (.codex/hooks.json, PreToolUse)
# return permissionDecision "deny", or exit 2 with the reason on stderr

# Claude Code, the reflex that saves sessions
Shift+Tab         # cycle plan mode and permission modes
Esc Esc           # rewind to an earlier checkpoint

The routing table we ended up hard-coding

After a few months of running both, the routing stopped being a judgement call. The split follows one question: can we describe the finished state in a paragraph? If yes, Codex. If the agent has to discover what the task is really about, Claude Code.

Here's the thing. We liked that table enough to put it in code. Maestro, the multi-agent mode in AITerm, routes every block of a plan by tier, and the defaults in MaestroRouting.swift are the table above written as rules rather than habits.

The 3-run, 15-point threshold exists so the routing never zigzags after one lucky run. A rule that flips every afternoon is worse than no rule at all.

The cross-review row deserves its own sentence. Handing one agent's diff to the other finds more real problems than any self-review, because the disagreement is never in the boilerplate. It is always on the decision that mattered.

Can you run both at the same time?

Yes, and past two agents it becomes a monitoring problem rather than a compute problem. A modern laptop runs four of these without noticing. What breaks first is your ability to know which one is waiting for an answer, and which one is about to hit its limit.

The mechanical part is git worktrees. The Codex desktop app now manages its own, in $CODEX_HOME/worktrees, keeping the 15 most recent by default. In a plain terminal, four commands cover both agents.

git worktree add ../proj-codex  -b feat/tests
git worktree add ../proj-claude -b feat/refactor

cd ../proj-codex  && codex     # scoped work
cd ../proj-claude && claude    # exploratory work

The quota side is less obvious, and building our gauges taught us why. The two agents expose usage in completely different ways. Claude Code's quota comes from Anthropic's usage endpoint, which the CLI itself calls with the same account, so a gauge that polls too eagerly can get both of them a 429. Ours polls at most every 5 minutes in the background, every 2 minutes when the quota page is open, every 15 seconds on an explicit refresh, and backs off after any 429.

Codex has no such endpoint. Its rate-limit data is written into the session logs under ~/.codex/sessions, so the number you read is only as fresh as your last Codex session. A Codex gauge can show 80% used for a window that already reset. Ours marks those windows as expired instead of pretending.

AITerm is a native macOS terminal built for piloting AI coding agents. It runs Claude Code and Codex in side-by-side panes, shows each session's live status, and puts both quota windows in one sidebar with alerts at 75% and 90%. Its Maestro mode routes deep work to Claude Code, scoped work to Codex, and always has the other agent review.

The setup we run on the AITerm repo

Two agents, two worktrees, one window. Claude Code sits in a pane on the feature branch, Codex in a second pane on tests or the localisation sweep, the dev server in a third, and the quota gauges for both accounts sit in the sidebar where we can see them without switching apps.

This is the part where our product enters. AITerm is 12 MB of native Swift, rendering through CoreText. The feature that changed the two-agent workflow for us is the quota view: Claude Code and Codex windows side by side, with a notification the first time either one crosses 75%, then 90%. When two agents meter usage in completely different ways, seeing both clocks in one place is what stops a refactor from dying at 70% done.

None of the routing above needs our app. Worktrees, short memory files, one task per session and clearing early work in any terminal, including yours. What the app removes is the alt-tabbing to find out who is waiting, which stops being a small thing somewhere around the third agent.

The short version: install both, keep one canonical memory file, route by whether you can describe the finished state, and let the other agent review. That's it.

Going further

Sources

1. Codex models, OpenAI docs, October 2026: GPT-6.1 Sol recommended for complex coding, GPT-6 Astra and Luna, GPT-5.5 retired from ChatGPT sign-in on October 14, 2026.
2. Codex pricing, OpenAI docs, October 2026: Free, Go $8, Plus $20, Pro $100/$200/$500, Business, 15 to 160 GPT-6.1 Sol messages per 5 hours on Plus, no five-hour limit on Pro.
3. Agent approvals and security, OpenAI Codex docs, 2026: sandbox modes, on-request, never and granular policies, untrusted retired, Auto preset, auto-review, network off in workspace-write.
4. Hooks, OpenAI Codex docs, 2026: twelve hook events, hooks.json and config.toml, PreToolUse deny, "a useful guardrail, not a complete enforcement boundary".
5. Subagents, OpenAI Codex docs, 2026: TOML agents in .codex/agents, built-in default, worker and explorer agents, per-agent model.
6. Memories and Git worktrees, OpenAI Codex docs, 2026: memories off by default, managed worktrees in $CODEX_HOME/worktrees, 15 kept by default.
7. openai/codex, GitHub: the Rust implementation of Codex CLI, AGENTS.md handling and the apply_patch primitive.
8. Model configuration, Anthropic Claude Code docs, October 2026: Opus 5.5 default on Pro, Max, Team and Enterprise, native 1M window, auto-compact near 967K tokens, opusplan.
9. Manage costs effectively, Anthropic Claude Code docs, 2026: about $13 per developer per active day, $150 to $250 per month, under $30 per active day for 90% of users.
10. Create custom subagents and Hooks reference, Anthropic Claude Code docs, 2026.
11. Plans and pricing, Anthropic, October 2026: Pro $20 or $17 annual, Max from $100, Team Standard and Premium seats, Claude Code not in the free plan.
12. Stack Overflow Developer Survey 2026, AI section: Claude Code (66%) and GitHub Copilot (59%) lead coding agents.
13. Terminal-Bench 2.1 leaderboard, BenchLM mirror, updated October 17, 2026: GPT-5.6 Sol 91.9%, Claude Opus 4.8 74.6%, no Opus 5.5 or GPT-6.1 Sol entry.
14. Claude Code vs Codex, Tech Insider, July 2026, citing Firecrawl (June 2026): one matched task at about 192,000 tokens and $2.50 versus 136,000 tokens and $2.04.