Implausible generation speed number in Assistant

Please see this Assistant thread as an example. 48,938 tokens in 197 seconds works out to 248 tokens per second, yet the UI shows 22 tokens per second, so this is confusingly off by an order of magnitude.

(It might be plausible that the 22 tok/s measures only the final output tokens, and disregards reasoning and tool call/response tokens. But it would be a little weird to measure this way over the total request duration, because the LLM is spending most of its time reasoning and doing tool calls.)

I would expect to see a number around 248.