Long AI chats: higher cost and weaker answers
In brief
According to Journal du Net, a single thank-you to a large model is cheap at global scale; choosing the model matters about thirty times more. The larger cost comes from resending the full history with every turn, and cited studies show reliability falls as context grows. The article recommends a coordinating conversation plus a separate thread per task.
Source : Journal du Net - « IA : ce n'est pas le "merci" qui coûte, c'est la conversation qui s'allonge »
What it changes for an SME
If you pay per token, each new message in a long thread typically resends much of the history, so cost grows faster than the number of turns. On consumer chat apps you may not see that bill, but the quality effect remains: longer threads are more likely to drop early instructions and mix versions, which costs staff time. Switching models in the middle does not transfer what the first model processed and triggers a full reread. Splitting a project into short, one-task conversations limits both cost and rework.
What you can do
- Start a new conversation for each distinct task instead of keeping one long thread for a whole project.
- Use one coordinating chat for the goal and framing, and do not produce the deliverables there.
- If you need a different model, open a fresh conversation rather than switching inside a long thread.
Let's talk about the time you could save.
In 20 minutes, we find where AI could save you time or money. Free, no commitment.