I believe that if you account for the fact that it is probably middle and high school students who use it the most (also much larger population than college) and the data would make more sense.
Replies
Do you know middle schoolers that are using OpenRouter in their code to switch between models easily? Because that's what the OP's graph shows.
aug 8, 2025, 1:54 pm • 2 0 • view
Can’t say I know any middle schoolers at all. I don’t even know what Roblox is besides it exists and is a thing.
aug 8, 2025, 2:04 pm • 0 0 • view
This ++ the dips before the plummet also map relatively well with different schools wrapping, with increased usage as the next batch of schools approaches finals, and then a collapse once MS/HS ends
aug 7, 2025, 5:44 pm • 19 0 • view
Reading too much into the data. If that were the case there would be spikes during an overall downward trend. And nonetheless I doubt we'd be able to separate them from each other so clearly. The rest of the dataset has a similar level of relative variability.
aug 7, 2025, 10:00 pm • 0 0 • view
In fact, the spikes throughout the dataset consistently last for 5 days. I would bet money the dips are simply just weekends.
aug 7, 2025, 10:01 pm • 0 0 • view
I disagree, but regardless it doesn't explain why usage on other models trends upwards, eg. openrouter.ai/anthropic
aug 7, 2025, 5:47 pm • 10 0 • view
It makes fine sense, bc until this week people serious about doing a summer project (or half assing their wfh job once their kids are home the the summer) turned to claude
aug 10, 2025, 11:35 pm • 0 0 • view
here's the overall trends (openrouter.ai/rankings?vie...) 4o-mini is the red at the bottom, seemed to have a weird usage spike around May, for whatever reason
aug 7, 2025, 8:17 pm • 9 0 • view
is it possible the new anthropic model just has increasing buzz/market share?
aug 7, 2025, 7:11 pm • 1 0 • view
i think i agree that this chart can’t be used to show openai’s business, like surely most chatgpt users are using it from the website and not through whatever tool or marketplace this website counts from, right?
aug 7, 2025, 7:15 pm • 2 0 • view
I have some technical knowledge but not a lot. Looking at the graph you have linked... it definitely looks like Sonnet 4 got launched end of May and processes more tokens in its normal operation. Seems plausible to me that people upgraded from 3.7 to 4 and created inflation in # of tokens.
aug 8, 2025, 5:26 am • 0 0 • view
Does "tokens processed" refer to input, output, or both? Is it normal that Anthropic is processing so many more tokens, when, as far as I'm aware, OpenAI has a greater user base. (I'm unfamiliar with openrouter so maybe its a quirk of that platform?)
aug 8, 2025, 5:34 am • 0 0 • view
To be clear, I do believe the core argument that AI is bunk and students are using it to cheat is accurate. I just disagree that this chart is supporting that argument.
aug 7, 2025, 5:49 pm • 11 0 • view
Do you under that all schools in the US don’t end on June 6? If it was what you represented, you see a gradual drop off from beginning of May to end of June
aug 10, 2025, 7:43 pm • 0 0 • view
Not really. Those schools go to mid-June or later, long after the dropoff in that graph. (Also I severely doubt that many middle school kids are using OpenRouter and making AI calls instead of using a model through a chat interface directly from the developer.)
aug 7, 2025, 5:44 pm • 0 0 • view
Came here to say this. No evidence of spread for range of college semester end dates; and most K-12 schools I know run much later than this, well into June. That sharp drop right at the beginning of June is just *too* abrupt. Great theory though!
aug 7, 2025, 5:47 pm • 1 0 • view