Skip to content
Comparison

Codex or Claude Code: which to choose — the ToTheMoon hosts' experience

OpenAI's Codex and Anthropic's Claude Code are agentic systems: they don't just answer, they do the work on your computer themselves — sorting through email, files and finances, writing and running code. The ToTheMoon hosts are convinced they are useful to more than just programmers, and they keep comparing the two. This page covers which tasks the hosts give each system, which models they run in them, where they hit the limits, and why they advise against putting all your tasks in one place.

Updated November 16, 2026 · ToTheMoon editors · from the hosts' experience, no editorial ranking
Point by point

Codex or Claude Code

What we compareCodexClaude Code
What the hosts use it forLife and business: hotel refunds, cameras without an API, a voice-recorder archive, expense tracking, watching over long computations.Development and automation; also analyzing the ToTheMoon episode archive and transcoding video on Amazon servers.
StrengthsChatGPT stands behind it: email and services are easier to connect, and Dots creates its tasks specifically in Codex.For development, per Alexander Volchek, “cooler than Codex”: its models Opus 5.5 and Fable on Max beat 6 Astra and 6.1 Sol.
Which models to useFor finance, GPT-6.1 Sol or GPT-6 Astra on Extra High or Max, Alexander advises; Codex has no Pro mode.Opus 5.5 or Fable 5.1 on Max; Alexander recently moved a task with too much money at stake from Opus 5.5 to Fable 5.2.
Limits and plansPaid tokens with auto top-up can be enabled; a reset marathon ran from October 5; Ultra Fast ate Alexander's whole Pro limit.Claude's limit runs in five-hour windows and Anthropic is stingy with resets; on Fable 5.1, subscription tokens once ran out in two hours.
Files, email and memorySearches all ChatGPT chats without browser access; together with ChatGPT and Dots, it remembers Alexander's everyday life.In September it wouldn't take files over twenty megabytes or work with the file system; Projects bring together repositories, environments and connections.
Weak spotsWorst of all at keeping up recurring tasks, and drifts from the goal: a couple of billion tokens went into one Linux server.Dictated tasks were lost after an interface update; developer Boris Cherny thinks it's time to rewrite Claude Code.
Risks and accessBanks — read-only; Alexander doesn't give Codex browser access: chats can be searched without it.With a $350 limit, Amazon showed $37,700 in charges for Claude Code's work; in the end, Amazon turned out to be at fault.

In short: what's the difference

Codex and Claude Code are built in a similar way. They are not a chat but a system that acts together with your computer, launches several agents on its own and can run a task for days: Alexander Volchek's longest task in Codex has been going for fifty-eight days. The main difference, if you listen to Alexander, is the ecosystem behind each. Codex is tied to OpenAI's ChatGPT, Work and the Dots assistant, while Claude Code runs Anthropic's models — Opus and Fable. Both, in his assessment, have pulled ahead of Cursor and Google Antigravity — precisely thanks to the model infrastructure.

Under Alexander's tenth rule, everyone should automate their life in one of these systems, not just developers. What holds people back, he observes, is not complexity but not understanding what to do there — which is why the podcast gives so many everyday examples. Alexander's own formula is this: Codex is for life, email and services, Claude Code is for development. The other hosts don't bet on a single horse: Ilnar Shafigullin and Anna Volchek deliberately work in both systems.

When the hosts choose Codex

Alexander's most vivid everyday stories involve Codex. After a trip around Europe, he gathered his bookings from email, found the hotels that had forgotten about breakfast credits, and kept up the correspondence until some of them refunded the money. He connected to a camera cloud that has no API, sorted fifteen to sixteen years of voice-recorder files into folders, and in a couple of minutes found 324 old contacts where Siri AI had found eight. And Codex identified a server computer he had bought with his personal card on its own, from his email, invoices and receipts.

Alexander Volchek's main argument for Codex is that ChatGPT stands behind it. Email and services are easiest to connect to it, the ChatGPT, Codex and Dots combination has learned his daily life over thousands of chats, Dots creates its tasks specifically in Codex, and Codex itself now searches ChatGPT chats without browser access. Ilnar Shafigullin trusts it to watch over computations that run on a server for five days or longer: every fifteen to thirty minutes it checks whether anything has crashed and, within what is permitted, restarts it.

When they choose Claude Code

For development, Alexander Volchek chooses Claude Code and says plainly that as a programming environment it is “cooler than Codex”. His argument is the models: in his experience, Opus 5.5 and Fable at the Max level outperform 6 Astra and 6.1 Sol, and Opus 5.5 is currently number one for him when it comes to mass development and automation. Fable impressed him even before the block: he launched it in Claude Code from his phone while sitting by the pool, described a big task by voice — and the model, knowing the infrastructure, saw it through to the end.

For financial accounting in Claude Code, Alexander recommends Max mode and one of two models — Opus 5.5 or Fable 5.1. And when Fable 5.2 came out, he moved his riskiest money task onto it from Opus 5.5 overnight — he didn't fully trust Opus 5.5. In September, Claude Code got Projects: a project receives a shared goal, and the tasks inside it run in parallel, each in its own branch. Alexander himself used Claude Code, partly together with ChatGPT, to go through the transcripts of more than 170 ToTheMoon episodes and pull out twenty ideas that had kept coming up for two years.

What has changed in recent months

The autumn was dominated by limits. Tibo from OpenAI, who is responsible for Codex, launched a marathon on October 5: every day brings either a useful update or a limit reset. The catch is in the details: pressing reset a day before the scheduled reset pushes the next one back by a week, and when the limit does run out, all your daily tasks stop. Ilnar Shafigullin, on the $200 plan, ran into a model-overload banner for the first time and, like Anna Volchek, started working in Claude Code in parallel. Anthropic is stingy with resets: Ilnar attributes this to cautious compute purchasing amid a sharp rise in the number of users.

The products themselves have changed too. OpenAI is merging ChatGPT with Work, and Codex has learned to search ChatGPT chats and suggests the next prompt in a pale gray font. Anthropic released Opus 5.5 and Fable 5.2, and Boris Cherny, one of the main creators of Claude Code, declared the tool outdated: according to Boris, it needs to be erased and written anew. That is why the hosts advise dividing your tasks between the two systems in advance: if an account gets blocked or the reset marathon ends, your work won't come to a complete halt.

Questions

Codex or Claude Code: in brief

Is Codex or Claude Code better?

The ToTheMoon hosts don't declare a winner. Alexander Volchek splits the roles: he gives development to Claude Code on the Opus 5.5 and Fable models, and email, services and everyday tasks to Codex, since it is built on ChatGPT. Ilnar Shafigullin and Anna Volchek deliberately work in both systems: that makes it easier to get through running out of limits or an overload at one of the companies.

Can I use Codex and Claude Code without coding skills?

Yes, Alexander Volchek believes: under his tenth rule, both systems are for everyone. Instead of fiddling with settings, he suggests giving the system a goal in plain words — for example, asking it to sort through your email and find unnecessary charges. If the system asks no clarifying questions or produces nonsense, then in his experience it isn't suited to that task.

What plan do I need for Codex and Claude Code?

Alexander Volchek advises starting with at least an inexpensive paid version and a strong model: on the lightest one, automation seems pointless. He considers a subscription of one to two hundred dollars justified from the moment you start automating things, and suggests keeping two — at roughly $20 and $100–200. A cheap plan won't do for Fable: it once used up a two-hundred-dollar subscription, he says, in four to five hours.

Can I trust Codex and Claude Code with my bank accounts and servers?

The hosts advise limiting access. Connect banks read-only, and if that isn't possible, forward statements to your email or upload them once a week. Alexander Volchek doesn't give Codex browser permission: the system has no reason to follow his mouse all day long. And after the story in which, with a $350 limit, Amazon showed $37,700 in charges for Claude Code's work, he advises setting explicit limits for the agent, although it was Amazon that turned out to be at fault that time.

Where they compared

Episodes where the hosts compare

Other comparisons

More comparisons

All comparisons →