Notes on coding agents and the pages they leave behind, on moving between notes apps, and on the plain mechanics of keeping what you think. Written and published from a Sumibako vault.
Claude Code vs Codex in November 2026: prices, limits, and two opposite ideas of safe
Claude Code vs Codex in November 2026: what each costs from Free to $500, the five-hour limits both dropped, and why their safety defaults are opposites.
Prashant Shaw
Founder, Sumibako
8 min read
I read OpenAI's page about its Pro plans this morning and it said "Updated: 4 hours ago". It now has three Pro plans, at $100, $200 and $500, and a paragraph explaining that a reduction it emailed people about in September has been cancelled. Days earlier, Anthropic renamed its $200 plan. Comparing Claude Code and Codex is like comparing two trains by their timetables while both are moving.
So here is the comparison as of 3 November 2026, from both companies' own documentation and nothing else: what each costs, what each limits, where each runs, and the one design difference that matters more than the benchmarks. I use both, I sell neither, and my own product works with both, which I mention once near the end.
🥊 The short answer
Claude Code
Codex
Made by
Anthropic
OpenAI
Models
Anthropic's: Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5
OpenAI's: GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna
Cheapest way in
Claude Pro, $20 a month
Free, with GPT-6 Luna in the desktop app as it rolls out; the CLI and IDE from Plus at $20
CLAUDE.md, or AGENTS.md when there is no CLAUDE.md
AGENTS.md
The row to read twice is the one about trust. It is the only place the two are genuinely different products rather than the same product with different logos.
💸 What each costs
Anthropic's pricing page puts Claude Code in every paid plan and none of the free one: Pro at $20 a month ($17 billed yearly), Max at $100 for five times Pro, and Ultra at $200, which "includes 10 times the Pro plan's usage every week" and, for verified accounts, no five-hour limit at all. Usage is one pool: "Your activity across Claude on web, desktop, mobile, and Claude Code all draws from the same pool." I wrote up every plan and model separately.
OpenAI's Codex pricing starts lower and ends higher:
ChatGPT plan
Price
Codex in it
Free
$0
GPT-6 Luna in the desktop app, "subject to rollout"
Go
$8
The same
Plus
$20
Web, CLI, IDE extension and iOS; GPT-6.1 Sol and GPT-6 Luna
Pro 100, 200, 500
$100, $200, $500
More of everything; Ultrafast mode on Pro 500 only
Business
$25 a user a month, or $20 billed yearly
Plus, with admin controls and larger cloud machines
Two things on OpenAI's page have no equivalent on Anthropic's. First, actual numbers: Plus gets an estimated 15 to 160 local messages per five hours on GPT-6.1 Sol and 350 to 3,000 on GPT-6 Luna, with the honest caveat that "These estimates are not fixed message limits". Anthropic publishes only multiples. Second, the same headline move: "Pro plans currently have no five-hour limit." Both companies have decided their top customers should only meet a weekly wall. Neither says how high the wall is.
The Pro tiers article is worth reading for the drama alone. Pro 100 and Pro 500 now "include twice their previous Work and Codex allowance", and for Pro 200, "The reduction announced in our September 29 email is canceled." If you were on Pro 200 and spent October annoyed, you may stop.
The bigger change is a month away. OpenAI's usage limits article says: "Starting in December, Plus, Pro, and Business plans' usage allowance will be shared across ChatGPT and Codex". Usage will be measured in tokens, and "Instant chat will not draw from your usage." That is exactly how Claude already works, one pool for chat and code, and it has the same consequence: a long afternoon of chatting with a thinking model now comes out of the same allowance as your agent. OpenAI expects "most users won't need to wait for their usage to reset more often than they do now." The forum thread asking OpenAI to keep the two separate passed a hundred replies in its first two days, so not everyone shares the expectation.
📝NOTE
Codex also runs on an OpenAI API key, billed at API rates, with no cloud features. Claude Code does the same with an Anthropic Console key. That is the route for CI and shared machines on both sides, and the subscription is the route for a person.
🛡️ Two opposite ideas of safe
Both agents can run commands on your computer without asking. They disagree about what should stop them.
Claude Code trusts a second model. Since version 2.1.283, "auto mode is the built-in starting permission mode for interactive terminal and VS Code sessions." In auto mode, "a second model, the classifier, reviews actions instead of you", and only what it refuses comes back to you as a question. Claude Code has an OS-level sandbox too, built on the same kind of tools, Seatbelt on macOS and bubblewrap on Linux, but the sandboxing page is plain about it: "The sandbox is off by default." You turn it on with /sandbox.
Codex trusts a wall. Its sandbox page says "The default permissions mode applies sandboxing automatically", and the sandbox applies to every command the agent spawns, git and test runners included. Inside the wall it works without asking; at the wall it stops. The second-model idea exists here as well, as an option: Auto-review "replaces manual approval at the sandbox boundary with a separate reviewer agent", which you pick as Approve for me in the app or set in the CLI's config.
So each company ships the other's idea as a setting. Anthropic's default asks "is this action sensible?" and OpenAI's default asks "is this action inside the box?" The first catches a mistake that stays inside the project, like deleting the wrong directory; the second catches a mistake the model did not see as one, like a command that reaches the network or another folder. My advice, which is also both companies' advice in their small print: turn on the half you did not get by default. In Claude Code, run /sandbox once. In Codex, try Approve for me. Neither costs you anything but a minute.
🌍 Where each one runs
Both have stopped being terminal tools. Anthropic's platforms page lists the CLI, a desktop app, VS Code, JetBrains, the web, where a session "continues after you disconnect", and the Claude phone app, which can start cloud sessions, steer a running local one with Remote Control, or send a job to your desktop with Dispatch. "You can mix surfaces on the same project."
Codex lives inside the ChatGPT desktop app as well as its own CLI and an IDE extension that installs in VS Code, Cursor and Windsurf, with JetBrains and Xcode integrations of their own. Codex Cloud gives each task its own workspace that "can keep working while your computer is asleep", and Codex on mobile lets you approve a command on a connected Mac or PC from your phone.
The practical difference is small and specific. Codex's editor reach is wider, Xcode especially. Claude Code's phone story is richer. And both extensions install in the same VS Code window, so nothing stops you running one in each panel and letting them argue.
📄 Instructions, skills and the good news
Here the two have converged, which is the nicest thing that has happened to coding agents this year. Codex reads AGENTS.md. Claude Code reads CLAUDE.md, and since version 2.1.277 it reads AGENTS.md too, but "only when you have no CLAUDE.md in your working directory or above it." Both support skills from the same open standard, from different folders (.claude/skills/ and .agents/skills/), and both speak MCP.
That means a repository can be set up once for both. I wrote the details up in CLAUDE.md, AGENTS.md, MEMORY.md and skills; the short version is an AGENTS.md with the facts, a CLAUDE.md that imports it, and skills kept in one place and linked into the other. Switching agents then costs you nothing but the habit.
⚖️ Pick by how you work
1️⃣ You want to try an agent for free. Codex. Claude Code has no free tier; Codex has GPT-6 Luna on Free and Go, in the desktop app, as it rolls out.
2️⃣ You pay $20 and code a few hours a day. Either. Claude Pro gets you Claude Code with Opus 5.5 and Sonnet 5.5; ChatGPT Plus gets you Codex with GPT-6.1 Sol and a published estimate of how far it goes. Pick by which model you prefer to argue with.
3️⃣ You run long jobs and hate the five-hour wall. Claude Ultra at $200, or ChatGPT Pro from $100. Both remove the five-hour limit; OpenAI does it on its cheapest Pro tier.
4️⃣ You care most about what the agent cannot touch. Codex's defaults, or Claude Code with /sandbox on.
5️⃣ You work in Xcode. Codex, which has an integration there; Claude Code does not list one.
6️⃣ You want both. $40 a month gets you Claude Pro and ChatGPT Plus, two agents with different blind spots, and a repository that already reads the same AGENTS.md for each.
🗂️ Wherever you end up
The thing neither agent keeps for you is the work itself: the plan, what was decided, what was tried, where the job stopped. Each keeps its own session history in its own format, and neither reads the other's. If you switch agents mid-job, or use both, the handoff has to live somewhere they can both reach.
That is the gap I built Sumibako for: Claude Code and Codex both file a page into the same vault with one command, npx sumibako, and read it back in the next session, whichever agent that is. It works on the free plan. It does not run code or replace either agent's own memory; it is a notes app with an API, and it is happiest holding the page that says what Tuesday's agent did so Wednesday's can carry on.
Both companies will have changed their prices again by the time you finish reading. Mine has not changed since September, which is either stability or a lack of ambition.
