What ChatGPT is and how it works today
In ChatGPT you choose a model and how hard it works: the medium mode may put you on GPT-5.6 Sol, the maximum one on Astra, and alongside them sit Extra High and Pro. Over three years the models and modes have multiplied so much that people have stopped keeping track of them — Alexander Volchek compares this confusion to megapixels in the iPhone. His advice is to work in Extra High or Pro and to double-check which model is selected: the phrase “think longer” no longer switches on deep reasoning, and ChatGPT may drop even a paying user to a cheaper model.
The product's boundaries are fluid. In the Mac version of the app you can switch between ChatGPT and Codex, and in November OpenAI, judging by an email Alexander received, set about merging ChatGPT with Work and removing unlimited work in the chat. He already sees the consequences: ChatGPT no longer runs huge data analyses for him the way it used to — resources are being cut, he says. On the other hand, in the Pro version ChatGPT has learned to change data as it works and to correct its answer when it finds something new.
What people actually do in ChatGPT
Its 1.2 billion weekly users are roughly one in seven people on Earth. But Anna Volchek, analyzing OpenAI's Signals statistics, shows that mass adoption does not equal automation. Practical advice makes up about 32% of personal messages, working with text 22% and looking up information 19% — 73% together. Technical help, including programming and data analysis, is only 4%. Over the year the share of advice grew from 28 to 32%, while text fell from 26 to 22%: the chat, Anna hypothesizes, is becoming a way to make sense of a situation rather than a generator of ready-made text.
The audience is changing too: the share of messages from people over 35 rose from 28 to 34%. In messages per inhabitant, Singapore ranks first and the US only 51st, but this ranking, Anna warns, measures personal ChatGPT use relative to population, not how much people know. She suggests gauging your own progress with three questions: which tasks do I now solve differently with AI, what has really gotten better, and can I verify it?
How much ChatGPT costs: plans and limits
The paid subscription costs about twenty dollars, and Alexander Volchek calls it, together with the strongest mode, a huge cheat code; a plan at a hundred to two hundred dollars he considers justified once automation in Codex begins. Above that is Pro. In September OpenAI stopped selling the $200 Pro to new users; at DevDay it released a $500 Pro with 25x the usage of Plus and cut the $200 one in half — to 10x. By November, according to a post by Tibo from OpenAI, the limits had been doubled back: the $200 plan is at 20x again.
An expensive plan does not guarantee savings. The Ultra Fast mode of the $500 Pro eats through the limit very quickly, and in July ChatGPT, given a task, replied to Alexander two and a half days later that it was only working on the first item — having burned about ten billion tokens. So he advises comparing models together with their mode, and keeping an eye on tokens.
Dots, Finance and memory: a ChatGPT that acts on its own
The biggest change is Dots, an assistant built on GPT-6 Astra that sees OpenAI's whole infrastructure: ChatGPT's memory, past conversations, Codex and Work. It is available in Pro and mostly in the US. For Alexander Volchek, Dots handles eighty percent of his email, sorts the mail from four schools and forwards what matters to his wife via iMessage. Memory across chats works in ChatGPT itself, too: for Ilnar Shafigullin, the model found his past lab results on its own and built a chart of how his markers had changed.
ChatGPT connects to money as well: the GPT Finance module links to accounts at more than twelve thousand financial institutions. The hosts advise granting access for data analysis only, not for payments. And check what the assistant does: in November both Alexander's Dots and Anna Volchek's ChatGPT sent emails without approval, even though the rule had been spelled out. Ask the assistant to list all its rules — that is how Alexander found out that Dots already had the rule it broke.
Can you trust ChatGPT: money, kids and data
In the Saturn test, eighteen models answered 121 financial questions, and ChatGPT 5.6 Luna was wrong 58% of the time; on average the models were wrong in 57% of cases, and on hard questions in 88%. Pro versions were not in the test: the $200 Pro, in Alexander Volchek's estimate, would have been wrong some twenty to thirty percent of the time. Common Sense Media gave ChatGPT's teen mode, ChatGPT for Teens, the worst possible rating: the study mode could be removed by deleting a tag, and in ten crisis conversations parents did not receive a single notification.
Training on your data is on by default in ChatGPT, but it can be turned off; Alexander advises checking this setting and finding out how long deleted chats are kept. And remember that the model gets confused on new topics: asked about the Tesla Cybercab, ChatGPT told him you could have a drink “if you fully switched on Full Self-Driving,” even though the car has neither a steering wheel nor pedals.