The count here includes cached read tokens, which get included based on all prior input tokens at every tool call.
Cached input tokens are 10x cheaper than input tokens. If you hover over the token count, it should give you a breakdown by token type.