I run the viberank (claude code / ai coding) leaderboard. Just crunched all 29k days of usage data and some of it surprised me
A recent analysis of 29,000 days of usage data from the Viberank leaderboard reveals surprising insights into how users interact with Claude AI for coding tasks. The data shows that cache reads constitute an overwhelming 94.8% of all tokens processed, meaning Claude re-reads context extensively for every token it generates. Furthermore, a significant portion of users, particularly those on the $100 or $200 Max plans, are incurring API-equivalent usage costs exceeding $1000 per month, with 11% of users exceeding $200 in a single day. Notably, 90% of users exclusively utilize the Opus model, indicating its widespread adoption beyond special occasions. AI
IMPACT Reveals heavy reliance on context re-reading and widespread adoption of premium models among AI coding users.