Claude-Code is getting dumber

Unlike before (I started around version v1.0.4+), the model used to understand context deeply — each prompt I made was understood correctly.

Now, even when starting from a fresh conversation, it hallucinates frequently. For example, I asked Claude to create a settings file where I could assign a specific xxx model and then use that variable throughout the codebase. Instead of following the xxx model I specified, it hallucinates and provides different models entirely. The model’s attention seems to drift far from what was stated in the prompt.

In addition, even when addressing very simple issues, it often takes 5–8 turns to arrive at a working solution. During this process, it tends to introduce numerous subtle bugs by adding unnecessary workarounds that affect other parts of the codebase — without reviewing how these changes might impact other components. This behavior is a notable regression compared to earlier versions.

It feels like i am prompting with model gemini-2.5-flash