It's funny how quickly humans adjust to new technology. Only a few months ago Claude Code and other agents felt like magic, now it increasingly feels like browsing the internet in the late 90s on a dialup modem.
Firstly, Anthropic has been suffering from pretty terrible reliability problems . And looking at OpenRouter's data they are not alone (nb. OpenRouter's data is not conclusive, but I believe it does give a somewhat interesting overview of reliability). If you've been using coding agents, you'll know how flakey they can be, often getting stuck and requiring retries, a bit like your 56k's modem dropping in bad weather or someone wanting to do a call.
This isn't unsurprising, as much as some commentators believe AI is overhyped, because AI token usage is absolutely exploding. While the 'Big 3' AI companies (Google, Anthropic and OpenAI) don't publish public statistics, OpenRouter does:
Again, we must caveat this data strongly:
Firstly, OpenRouter almost certainly routes a tiny proportion of LLM requests vs the global market, meaning these statistics could be distorting trends
Secondly, Grok (especially) is "dumping" a lot of free LLM tokens on the market getting feedback on their models via OpenRouter, which is probably twisting these statistics
The fact that we're trying to understand a revolutionary shift in software development through the tiny window of OpenRouter's data is telling. Google, Anthropic, and OpenAI guard their usage statistics like state secrets. The only glimpse we get is from OpenRouter, which likely handles <1% of global LLM traffic, yet even this tiny sample shows a 50x increase.
Given that agentic coding workflows consume probably something on the order of 1000x the tokens that non-agentic 'chats' or most API calls do, so it is not surprising to see such a big increase.
This is no doubt putting absolutely enormous strain on the infrastructure behind the scenes, which reminds me a lot of the first days of broadband when the ISPs really struggled to handle peak time loads on their interconnects.
Tok/s is all we need?
... continue reading