Cline and Roo-cline are responsible for 79% of the token usage for the month on openrouter, at a combined 548B tokens. That’s nearly 4 times the total combined usage of the other apps on the top 10 (151B). Per the model rankings on openrouter, it seems like Anthropic’s likely bearing the brunt of the increase. These numbers are only part of the total picture, since it’s excluding things like GitHub Copilot, Cursor, Windsurf, etc etc.
For those unfamiliar, Cline is an “Autonomous coding agent right in your ID”, open source and available for free on GitHub. It’s very similar to GitHub edits, but does a much better job of sequential reasoning (in my experience, at least). You give it an API key and your code base, and it handles putting together the instruction prompt with the relevant portions of your code. This tends to work better with smaller codebases that dont bump up against the context window, but there’s nothing stopping anyone from using it with a much larger codebase. And from the numbers, it’s seems like that’s happening quite a bit.
This means that these agentic coding apps are now responsible for monthly token usage that has to be orders of magnitude larger than the numbers that were getting hit prior to their release. (Maybe I’ll see if I can find historical openrouter data to back this up)
So my question is - how well are OpenAI and Anthropic handling this, and can they continue to scale at the pace that these tools are growing in popularity? Obviously both companies offer models on GitHub Copilot, and cursor, Sourcegraph Cody, etc have been around for a while, so it’s not like Cline popped up overnight, but still - but given the economical and supply chain challenges with scaling AI data centers, there’s no way this isn’t putting pressure on their poor ops teams.
1 comment
[ 4.7 ms ] story [ 12.8 ms ] threadFor those unfamiliar, Cline is an “Autonomous coding agent right in your ID”, open source and available for free on GitHub. It’s very similar to GitHub edits, but does a much better job of sequential reasoning (in my experience, at least). You give it an API key and your code base, and it handles putting together the instruction prompt with the relevant portions of your code. This tends to work better with smaller codebases that dont bump up against the context window, but there’s nothing stopping anyone from using it with a much larger codebase. And from the numbers, it’s seems like that’s happening quite a bit.
This means that these agentic coding apps are now responsible for monthly token usage that has to be orders of magnitude larger than the numbers that were getting hit prior to their release. (Maybe I’ll see if I can find historical openrouter data to back this up)
So my question is - how well are OpenAI and Anthropic handling this, and can they continue to scale at the pace that these tools are growing in popularity? Obviously both companies offer models on GitHub Copilot, and cursor, Sourcegraph Cody, etc have been around for a while, so it’s not like Cline popped up overnight, but still - but given the economical and supply chain challenges with scaling AI data centers, there’s no way this isn’t putting pressure on their poor ops teams.