Time to Switch From ChatGPT? I Tested Claude Duet on My Own Tasks
Is it worth switching from ChatGPT Dots to Claude Duet — and what really decides the choice of a single-window AI assistant if the models hardly matter anymore?
Claude Duet, launched a week ago for Pro subscribers in the US through a waitlist, appealed to Alexander Volchek much less than ChatGPT Dots: it can't work with sessions already created in Claude and acts only in tasks it sets up itself, and it also spends shared Claude tokens — once they run out, the assistant stops. Dots, meanwhile, has taken over eighty percent of his emails, the school mail it now forwards to Alexander's wife via iMessage on its own, and bank alerts — although it too is available only on Pro, almost only in the US, and listens to the infrastructure partially: a mess builds up in the single window, and it may remind you about a charge only once. What keeps such systems from being rolled out to everyone is cost: an active subscriber is a loss for the companies, and a billion active users would cost trillions. Choosing a model, meanwhile, has become a search for a needle in a haystack: Fable 5.1 is five to ten times more expensive than Opus 5.5 but may spend fewer tokens, and OpenAI's credits don't add up even in its own emails. Working through Dots, Alexander stopped thinking about models — and advises choosing one main system and, if you have the money, one extra.
What to watch for
Key takeaways
Claude Duet launched a week ago in a preview mode: only for Pro subscribers, only in the US and through a waitlist. Alexander got access four or five days later, and in that time the internet had already declared ChatGPT Dots dead. This, he says, is the same hype market where ChatGPT's death has been predicted for two or three years and every week brings a model "a thousand times cheaper."
People still don't properly use Codex and Claude Code because they don't know what to do there — which is why the channel gives so many examples. The same goes for Dots: people open it like the familiar ChatGPT chat, ask a question and look for information, without understanding what the system rests on and how it will develop.
Claude Duet has a separate app and Dots doesn't, and Alexander thinks ChatGPT should release one. At the same time, the companies already have too many entry points: Anthropic is closing the separate Claude Designs app and on September 16 merged Cowork and chat.
Dots is available only on Pro and mostly in the US. For a friend of Alexander's with a Singapore SIM card it disappears and reappears as soon as he connects to the home Wi-Fi; his sister on Pro never got it at launch, and in Europe it isn't available and isn't promised because of European Union rules.
Elon Musk said SpaceX could be worth one hundred trillion dollars. Some will call it pure hype, but Alexander recalls that he himself has said many times that a serious AI provider over a three-to-ten-year horizon — OpenAI or Anthropic — should be worth more than ten trillion dollars, unless huge competition appears on the market.
A setup where one agent looks for news, another updates data, a third does design and a sixth sends out sales information is, according to Alexander, absolute nonsense today. A project's task is worth writing out in full in one session; how many agents it uses shouldn't concern you, and claims of thousands of agents are marketing.
Today, when you set a task, you run into not the quality of the prompt but the quality of the model and your own vision: can you picture in advance what a page, a website or a report form should be. The model will give you very different options — not worse or better, but different.
An ordinary chat session has the model's limitations, above all the context window. Grok Bot, ChatGPT Dots and Claude Duet essentially have none: the huge correspondence is at once one window and many different windows, a destructured memory.
Alexander already sends eighty percent of his emails through Dots. School mail, which his wife used to forward, Dots now sorts itself and sends her messages via iMessage; bank alerts no longer come by email, and he opens his banking app less and less, because OpenAI is integrated, he says, with about ten thousand financial organizations.
Dots and Duet aren't rolled out to everyone because they are heavy and expensive infrastructure. A subscriber who squeezes a plan to the maximum is a serious loss for the company, and only a few do. If the whole billion users started doing it, costs would run into the trillions, and the companies couldn't handle such a load.
The main reason Alexander didn't like Claude Duet: it can't work with sessions already created in Claude. Duet read his tasks but replied that it can act only in those it creates and manages itself. "Anthropic has, for the first time, released such an anticipated product so poorly," says Alexander.
A single-window system mixes everything up: contract requests, a charge warning, a flight, a question about 3D printing. It's like managing projects through Slack or Telegram. Dots is supposed to listen to your infrastructure constantly — memory, new sessions, Codex — but it listens partially and may remind you about a five-hundred-dollar charge once and then go quiet.
Even model developers now say the context window doesn't always need to be shortened. Alexander's longest task in Codex has been running for fifty-eight days, his longest continuous run lasted eighty-eight hours, and in total he has used 64.7 billion tokens; at DevDay an OpenAI development lead talked about two-month sessions.
Claude suggested that Alexander translate his website's texts with a cheaper model, but nobody today knows what comes out cheaper and better — an expensive model from the start or a cheap one with refinement. Fable 5.1 is five to ten times more expensive than Opus 5.5 but may spend fewer tokens; on top of that come modes from low to max and Ultracode.
ChatGPT gave Alexander sixty thousand credits, which it valued at about two thousand dollars, and a week later OpenAI sent about twelve thousand credits as compensation for a mistake on Pro plans — "roughly a thousand dollars." Nobody understands how tokens are spent on the plans or how much a specific task will take.
Claude Duet eats your Claude tokens, and if they run out, you're "dead." In Alexander's view, this shows that Anthropic isn't betting on an everyday assistant for an ordinary person: its Duet examples are QuickBooks, Stripe, HubSpot, Salesforce, DocuSign, Canva and Shopify — in other words, running a business.
ChatGPT Dots is the ordinary person's story, and an ordinary messenger doesn't run out of limits: that only happened with SMS, which Alexander counted in 2001 on his Nokia 8310. But Dots will get limits too once it starts creating sessions in Work or Codex.
Working through Dots, Alexander noticed that he had stopped choosing the model — even when solving a non-trivial task of 3D-printing a latch for a basketball hoop — and switches on Pro mode less often. These systems are an entry point into infrastructure, and comparing small details becomes pointless: he buys Apple because he chooses Apple's infrastructure as a whole.
Alexander calls claims such as Moonshot's Kimi K3 being "at the level of Fable and Astra" absurd: Anthropic and OpenAI have even more powerful models. His rule is to choose one main system, use all the use cases in it and keep a second as an extra if you have the money; a third and fourth are hard to use.
What this episode is about
Claude Duet vs ChatGPT Dots: Alexander Volchek's Hands-On Comparison
A solo ToTheMoon episode in which Alexander Volchek takes apart Claude Duet — Anthropic's new system, launched a week ago for Pro subscribers in the US through a waitlist. He got access four or five days later and tried to continue his current work in Duet, and in that time the internet had already declared ChatGPT Dots dead. Alexander calls this the hype market and compares not only Duet and Dots but also two similar systems — Grok Bot and Meta Muse.
All four give you one window with a messenger interface instead of a chat with tasks and projects. Duet has its own app, which Alexander considers a good move; Dots is available only on Pro and almost only in the US — for a friend with a Singapore SIM card it disappears, his sister on Pro didn't get it, and in Europe it isn't available because of European Union rules. He calls working with fifty separate agents nonsense: a task is worth giving in full to one session, and quality today depends not on the prompt but on the model and your vision of the result.
Dots, Duet and Grok Bot essentially have no context window, and a single entry point can be a chat, a voice, a speaker or a screen in the car. Alexander already sends eighty percent of his emails through Dots: Dots sorts the school mail itself and sends his wife messages via iMessage, and bank alerts no longer come by email — OpenAI is integrated, he says, with about ten thousand financial organizations. Comparing this with DeepSeek or national models, he believes, is pointless.
He liked Duet much less: it can't work with sessions already created in Claude and acts only in those it sets up itself. What keeps such systems from being rolled out to everyone is cost — an active subscriber is a loss for the company — and the single window breeds a mess: contracts, charges, flights and 3D printing get mixed up in one window, and Dots may remind you of something important only once.
A separate part of the episode is about choosing a model and money. Alexander's longest task in Codex has been running for fifty-eight days, and in total he has used 64.7 billion tokens; Fable 5.1 is five to ten times more expensive than Opus 5.5 but may spend fewer tokens, and OpenAI's credits don't add up even in its own emails. Duet eats Claude tokens, and when they run out the assistant stops — which, in Alexander's view, shows Anthropic's bet on business, whereas Dots is the story of an ordinary person whose messenger never runs out.
Alexander's main conclusion: choosing between such systems is no longer a choice of model but a choice of infrastructure, like choosing Apple as a whole rather than by specs. Working through Dots, he stopped thinking about models, calls claims like "Kimi K3 is at the level of Fable and Astra" absurd, and advises choosing one main system and, if you have the money, one extra.
The episode presents single-window assistants not as yet another model but as an entry point into a company's infrastructure: the system that wins is the one built into your email, bank, tasks and sessions, not the one that is stronger in a single test.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 93 segments: 93 identified, 0 mixed, 0 probable, and 0 unresolved.
Read transcript on a separate page
Loading…