Artificial intelligence systems are no longer just for conversational responses: agents, capable of autonomously performing a series of tasks, are changing how we use them. In an updated guide, the author advises choosing between ChatGPT and Claude for real work, while being mindful of the permissions granted to the AI.
For the author, the change lies in the very nature of usage. An agent combines the “brain” of a language model, i.e., an LLM, with tools that allow it to plan and act. It can thus perform in one go the equivalent of several hours of human work. To illustrate these advances, the author asked GPT-5 in 2025 to create a brutalist city-building game, and then repeated the exercise with GPT-5.6 Sol in Codex, less than a year later.
For sensitive advice, such as a second medical or legal opinion, he recommends the most advanced models, Opus and Fable from Claude, or GPT-5.6 Sol from ChatGPT, set to the “High” reasoning level. According to the source, these models have lower error rates, but using them comes at a cost.
Two ways to give a computer to the AI
The simplest way is to use the virtual computers provided by the companies. At OpenAI, the mode is called Work; at Anthropic, Cowork. The user chooses the model, then the applications the agent can connect to, such as email. The author thus asked both systems to prepare a seminar: after about ten minutes, both had produced materials and an email, a task that would have taken two hours according to him.
This test also revealed a difference in settings. ChatGPT sent an email directly because the author had given it permission, while Claude only prepared a draft. The author recommends keeping pre-approval, which is the default setting, as long as you are not familiar with the system’s errors. He also warns against prompt injection, a malicious text hidden in a message or web page intended to hijack the agent.
The most powerful solution remains granting access to your own computer, via the ChatGPT or Claude applications. Codex and Claude Code modes show the details of the work: files modified, commands executed, tests performed. The author thus had his upcoming book, scheduled for release in October, reviewed by GPT-5.6 Sol in Codex. The agent worked for 30 minutes, checked 195 references, and produced notes that, according to the author, were free of errors, though he notes he had to reject some comments deemed too nitpicky.
According to the source, Google currently does not have a comparable top-tier model. However, the author cites Gemini Notebook, formerly NotebookLM, for research across many sources, and Gemini Omni, an LLM capable of seeing and directly editing video. For voice, he mentions GPT-Live, ChatGPT’s voice mode, which converses with the pauses and interruptions of a real conversation.
The author’s final advice is brief: choose Claude or ChatGPT, pay for the subscription starting at $20 per month, and assign a real task to an agent, then request corrections as you would from a person.
