What Codex is
OpenAI's lineup currently consists of ChatGPT, Codex, Work and Dots. Work is meant for documents, email and presentations without code, while programming stays with Codex. The boundaries are fluid, though: in the Mac version of the ChatGPT app you can switch between the chat and Codex. Above it all sits Dots — a single window that sees your projects in Codex and creates tasks right there. That is why Alexander Volchek advises learning Codex even if you are far removed from software development.
As in ChatGPT, in Codex you choose a model and its effort level — Medium, High, Max or Ultra Code, but without the familiar Pro. The same name, Alexander warns, guarantees nothing: in his experience, the Extra High mode of 5.6 Sol answers differently in the chat and in Codex. Codex launches several agents on its own, suggests your next request in faint grey text and can now search across all your ChatGPT chats.
What Codex already does for an ordinary person
The most vivid examples come from Alexander Volchek. After a trip around Europe, Codex gathered his bookings from Booking, Expedia, Marriott, Amex Travel and Chase Travel in his inbox, found the hotels that had not applied his breakfast credits, wrote to them itself and kept up the correspondence until some of them refunded the money. It also helped him reassign a French speeding ticket to a friend. And when Siri AI found five, and then eight, old contacts, Codex tracked down 324 in a couple of minutes.
Footage from Alexander's twenty-plus cameras sits in Amazon's cloud with no API, yet Codex connected to it on its own and is downloading the recordings. His iPhone voice-recorder archive — tens of gigabytes covering fifteen or sixteen years — it sorted into folders, spinning up an inexpensive local model for transcription. The host replaced the app for his heart sensor with Codex commands: turn the sensor on for the night, export the data, put together a report, send it to the doctor.
Money is a story of its own. In the episode about OpenAI's assistants, Alexander showed how the system broke down his July car expenses in two seconds, suggested on its own that Uber and Lyft rides be included, and produced a monthly cash flow and P&L. When Alexander bought a server computer for the company with his personal card and briefly asked Codex to account for it, Codex found on its own, in his email, invoices and receipts, what the purchase was. Hence the host's forecast: bookkeeping at 99.999% of legal entities will move to Codex, Claude Code, Cursor and Grok.
How to set Codex a task: the hosts' advice
Alexander Volchek suggests starting with a paid plan and a strong model: a friend of his complained about “dumb ChatGPT Codex” while running the lightest one. For financial bookkeeping the host recommends GPT-6 Astra or GPT-6.1 Sol at Extra High or Max. After that, set a goal instead of fiddling with settings like agents.md: for example, ask it to go through your email and find out whether you have been overcharged. If the system does not ask clarifying questions or produces “complete gibberish,” then, in Alexander's view, it is not fit for that job.
He walks through a big task using the example of a dossier on Greg Brockman: download seven years of his YouTube interviews, transcribe them with a local Whisper, go through at least a thousand sources and suggest questions. Work like this deserves a complete description — the result, the sources, a format such as an HTML report — and it should start with a general audit and be split into five to ten stages, with GPT-6 Astra Max set for it and time set aside for corrections along the way. A ToTheMoon episode transcript from besolid.com can also be handed to Codex as instructions.
The rest of the advice is about access. Connect your email not to clean out the inbox, but so that Codex goes through it every day. Give banks read-only access; if that is not possible, forward the statements to yourself or upload them once a week. Alexander flatly refuses to grant browser permissions. Ilnar Shafigullin has Codex check multi-day computations every fifteen to thirty minutes and restart whatever has crashed within the permitted limits, and to keep the work going with the laptop closed, Codex is set up on your own server and connected over SSH.
Limitations, usage limits and debates
Recurring tasks, Alexander Volchek observes, are what Codex handles worst: clarify something in the chat and it may drop a whole block of what it was doing. Anna Volchek sets such chats up again every one to two weeks, and Ilnar's hourly task ran properly only for the first twenty to thirty hours. Alexander also notices that Codex drifts away from the goal: recently a couple of billion tokens went into configuring a single Linux server, and the encryption was never finished.
The second pain point is limits. When they run out, all your daily tasks stall, and someone on the $20 plan, Alexander reckons, risks waking up to a “You've hit your limit” notification. In Codex you can enable paid tokens with auto top-up, say a hundred dollars at a time, and the Ultra Fast mode of the $500 Pro plan burned through one hundred percent of the host's Pro limit. Since October 5, Tibo from OpenAI, who is responsible for Codex, has been running a marathon: every day brings either a useful update or a limit reset.
The main debate is Codex versus Claude Code. In Alexander's experience, email and services are easier to connect to Codex, since ChatGPT stands behind it, and he rates ChatGPT above Anthropic's system. But for development Claude Code, in his words, is “better than Codex”: at the Max level he rates the Opus 5.5 and Fable models above 6 Astra and 6.1 Sol. Cursor and Google Antigravity, in his estimate, lag behind both. Hence the advice from the episode about limits: split your tasks between Codex and Claude Code in advance — in case your account gets blocked or the reset marathon ends.