So this happens rarely, but when it happens it is very obvious.
Claude Opus in Antigravity sometimes gets stuck in the thinking state for way longer than normal. The weird part is that the extra thinking doesn’t lead to a better answer. It usually leads to the opposite.
When Opus behaves normally, it is fast, direct, and usually very strong. It doesn’t sit there “thinking” forever before doing simple work. But in these stuck-thinking cases, the final result is often bad. Sometimes it hallucinates, sometimes it seems to lose the actual task, and sometimes it responds like it spent the whole time going in circles instead of solving anything.
This feels more like a bug in the Antigravity model/tool loop than normal model reasoning.
I don’t know if this is related to thinking mode, tool calls, provider routing, or the way Antigravity handles Claude responses. But the pattern is consistent enough that “more thinking” is not the right explanation. More thinking should improve the answer. Here it usually predicts a worse one.