This reminds me of something that came up recently. Copilot started hallicinating quite a bit more than usual in Copilot reviews. That made me think about the cost of operarion. As they burn money like this, I won’t be surprised if they start decreasing inference quality to decrease cost per user. Which also means people relying on certain model behaviour for tasks could get nasty surprises. Especially within automation workflows where model outputs aren’t being reviewed.