Want to bet NO but I strongly suspect OpenAI is lying about this and will lie again

I did not bet here but Sam Altman seems very much the type to happily lump API users with UI users to make line go up.

Hang on a sec

DAU/MAU of 36% is very very high

And andreesen horowitz is a very credible source

kudos for @256 for pointing this out again haha

@ian I'm starting to think this question may resolve NO, despite the 900M WAU number being true

The WAU/MAU number may really be above 0.9

Like you can model it this way

with:

(where N is the number of active users)

Then for the relationship between DAU and MAU you get this:

Then with the 36% DAU/MAU number from a1z you can solve for p:

Which basically gives you p=0.36 because the denominator is very close to 1

and then you plug it into this:

Which spits out WAU/MAU = 0.956

And so MAU = WAU / 0.956

If we take WAU = 900M then that gives us MAU = 941M

I'm fairly convinced that chatgpt does not have 1B MAU now, but if anyone's good at maths maybe they could give some input here?

@Magnify I am so sorry and you were right all along 🙏 🙏 🙏

like maybe modelling it as each user being a Bernouilli random trial each day isn't right?

🙇

chatgpt don't need no owl

the DAU/MAU number from a16z needs to be off by a lot for this model to start predicting MAU >= 1B

@Dulaman correct me if I'm wrong, but doesn't your analysis assume that each user uses the site with an equal frequency? Like imagine we had three types of users: casual, moderate and hardcore. Hardcore users use ChatGPT every day, moderate users use it once per week on a random day, and casual users use it once per month on a random day. Of the 36% figure, if the split is (for example)

20% hardcore

15% moderate

1% casual

(the real composition would probably be based on a binomial distribution)

Then the ratio of hardcore : moderate : casual would be 20 : 60 : 30. The WAU:MAU ratio would then be (20 + 60 + .25*30)/110 =~ 80%.

@ItsMe I think there's different ways of writing this. Some of them average out to a model like DAU = N * p, and some do not

@ItsMe like if each user has a different p value, and this is written as a random variable P over the set of users, then the model averages out to DAU = N * mean(P)

If the users aren't following Bernoulli random trials with fixed p values per user, then it breaks down

More of a fun fact than an argument for resolution: Duolingo has 47m DAUs and 130m MAUs. That is a very low correlation, especially for an app that so heavily encourages streaks. I wish they reported weekly active users too.
(https://investors.duolingo.com/news-releases/news-release-details/duolingo-surpasses-50-million-daily-active-users-grows-dau-36)
(https://intelpoint.co/insights/in-four-years-duolingos-daily-active-users-nearly-quadrupled-monthly-active-users-doubled-and-paid-subscribers-increased-tenfold/)

@256 that owl has been slacking

@256 chatgpt doesn't even have an owl to tell you to come back

@Dulaman it is a similar retention to that of ChatGPT
> After becoming the fastest product to ever hit 100M WAUs (in 2023), ChatGPT has sustained massive consumer adoption with an estimated 800 – 900M WAUs across platforms. Gemini is at 34% of ChatGPT’s scale on web, and 40% on mobile. ChatGPT is also dominant when it comes to engagement and retention. Per Yipit, ChatGPT’s DAU/MAU of 36% is nearly double Gemini’s 21%. And, month 12 desktop user retention of 50% is similarly double Gemini’s 25%.
(https://a16z.com/state-of-consumer-ai-2025-product-hits-misses-and-whats-next/)

@256 is there a mathematical model to relate typical DAU WAU MAU numbers?

like if you know the DAU and WAU then you can probably predict the MAU pretty closely

@Dulaman not for free, no.

@256 that a16z link swung me to NO

That DAU/MAU number (when plugged into the maths) is a significant update to my mental model of what chatgpt's user behavior stats are like