Usage meter fluctuates quickly and inaccurate

After the update with including the usage meter, I have experienced lower usage with models like GPT 5.2 (Instant) compared to the credits system. I do not paste in heavy text or files - moreso medium size like a page from a Github Readme and 1-2x Web Search set - this eats up usage quite a bit. The meter appears to be inaccurate because when I am continuing a thread, I cannot send any messages although the meter is 38%. Even after waiting for the 30 second interval