#A study of LLM usage across 100 trillion tokens
---
date:
· theme: eval
· words: 162
---
A study on LLM usage was just published, based on 100 trillion tokens. It covers all the main models: Gemini, ChatGPT, Claude, Grok, Llama, DeepSeek...
A few interesting points:
- Open source accounts for 30% of total volume. Closed models no longer have a monopoly: Llama and its peers take a third of the market.
- Coding is the undisputed number one use case. In that vertical, developers favor the Claude family.
- Reasoning matters: 50% of traffic goes to reasoning models, especially since o1 came out a year ago.
- The Cinderella effect: loyalty correlates with solving a new pain point. The first cohorts of users who saw a model fix their specific problem stay loyal far longer than the ones who come after.
- The roleplay surprise: roleplay usage is huge, especially on open-source models, and even makes up 80% of DeepSeek’s usage (!)
We built the most expensive infrastructure in human history to play with elves in a dungeon 🤷♂️
The full study: openrouter.ai/state-of-ai
