I hardly use ChatGPT or Claude anymore

ChatGPT was a reflex for me for a long time. Over the past three months, I have barely opened it. My use of AI continues, with a fairly clear shift toward tools that spend less time replying in a tab and more time working in my environment.

ChatGPT has become a reflex for the masses

When people talk about consumer AI, ChatGPT is still often the first name that comes up. On February 27, 2026, OpenAI stated that more than 900 million people were using ChatGPT every week. At that scale, ChatGPT has moved beyond being a tool people try out of curiosity. For millions of people, using it has become an everyday action.

An email has you stuck, so you ask ChatGPT. A subject seems unclear, so you ask it for an explanation. You might turn to it for rewording, an outline or a code fix. Some people also use it to prepare a reply before sending it.

Since 2020, conversational interfaces have proliferated. Claude, Gemini, Perplexity, Le Chat and Copilot have brought the same approach into many products. Assistants are also appearing directly in software people already use every day.

The current situation can be summed up in one sentence. Talking to an AI has become an almost automatic action.

My usage has moved out of the chat tab

Yet over the past three months, I have realized that ChatGPT and Claude have almost disappeared from my daily life. The need for AI is still there. My reflexes have simply moved elsewhere.

With a conventional conversational AI, you write a sentence and the model replies. It is useful for clarifying, rewording or getting an idea unstuck. Today, I expect something else. I want to give an AI an instruction and see it act. That is why I visit chatgpt.com less often and go straight to applications installed on my computer.

It started with Copilot, then Claude Code in the CLI. For the first time, I used an AI that could act directly on my computer. Going back and forth between chatgpt.com, claude.com and my files lost its appeal. You quickly feel the difference. The AI works in the same environment as I do.

The rapid success of agentic frameworks such as OpenClaw and Hermes shows that many people were waiting for this type of application. These tools are still mostly used by technical people, the nerds on X or Reddit. When I talk about them with people around me, many nontechnical people struggle to believe that an AI agent can actually act like this.

With Copilot, then Claude Code, and recently Codex, the separation between me and the AI has narrowed. Tools confined to the terminal are gradually making room for more accessible desktop applications. The agents retain the ability to interact directly with our work environment.

Codex follows an approach similar to ChatGPT. I talk to it, it replies, I correct it, it adjusts. What makes it different is its integration with my computer. It can read projects, modify files, run commands or open a browser. It can also check results, handle documents and use connectors such as Gmail or Drive.

It is so connected that I barely open my inbox. I simply ask Codex to read my emails and prepare draft replies. Yes, really…

Agents are arriving on the desktop

Everyone knows ChatGPT. Far fewer people already know that there are AI agents capable of acting on their computer.

Yet these tools are becoming easy to try. You download them, launch them, then watch what they can do. I feel that this shift toward more accessible uses is going to matter enormously.

Many people still use ChatGPT to ask for a list of places to visit on vacation. Meanwhile, I use Codex to analyze my professional WhatsApp conversations and track follow-ups in a Google Sheets spreadsheet on my Google Drive.

These tools also accelerate failure

I love the era we live in, because I have never been able to create so quickly. Still, I wondered whether these tools were reducing my intellectual stimulation. After three months, my conclusion points the other way.

I accomplish more tasks with less friction. Creating a project becomes faster. My projects launch faster, even when they fail more often. Ultimately, I can launch several in a shorter period of time.

Before, I might launch 5 projects in a year, with 1 succeeding. Today, I can launch 5 per month, with 1 succeeding. Success comes faster. So does failure. This pace calls for new mental flexibility and new digital hygiene. It can become addictive, sometimes toxic.

Trying agents with safeguards

If your use is still limited to chatting with ChatGPT in a tab, you are still at the beginning of what these tools make possible. Codex, Claude Code and other agents are worth trying, even briefly.

The strength of these tools comes mainly from their connection to entire ecosystems. They can access local files, the browser or Gmail. Drive, Slack, the calendar, scripts and plugins add still more entry points.

There is no magic formula for using agents well. The most solid approach is to experiment, correct, rerun and provide the right context. The perfect prompt becomes secondary when getting started. You test, then adjust;)

Backfires happen too. With great power comes great responsibility. These tools are still software with permissions, potential errors and real effects on data. Stories of important folders deleted by AI or security vulnerabilities introduced into personal content are already circulating widely.

Going further

To explore this subject further, I also discuss it in my book Le super pouvoir de l’IA, available online and in stores.

Sources

• OpenAI states that more than 900 million people use ChatGPT every week
• The open source Codex CLI repo lets you examine the tool's code