By skyforbes By measuring message count and 'Usage limit remaining', the last few measurements of what 100% is in terms of messages has come out as 250.0, to the limit of accuracy (usage remaining is reported as a percentage with one decimal point) It was higher initially. According to my measurements, token usage is not correlated at all to usage consumption, which seems hard to believe, but it's what I see.
I have model routing on, but it looks biased to the premium model, about 80% of my messages in this typical session are Pro.
This is using oauth authentication of a personal gmail account with a Google "premium" developer plan, which seems to have the same quota as AI Pro.
250 surprised me. My subjective feel with 2.5 quota was around 120 to 130, but one was never sure.
A "round" of messages for what I consider a medium sized task (say, a feature addition requiring a schema changes) is about 20 messages roughly. So the paid plan is not enough for a day of AI assisted work, but for what you get compared with competitor plans, I think the value is correct. We just miss more tiers, particularly if you just want to use only coding assistance.