OpenRouter uses ClickHouse Cloud as its single store behind customer analytics, public rankings, internal AI agents, and routing. Since adopting it in November 2024, monthly requests grew 33x to 19.7 billion, with 110+ trillion tokens served weekly and a 134-billion-row generation log. The same data powers customer dashboards, public model rankings, internal Slack-based AI agents, and real-time routing decisions across 500+ models. The newest addition, the Jev Router from TypeSafe AI, available since October 2026, lets OpenRouter pick the model itself based on real usage data, not just the provider.
Table of contents
Growing up with ClickHouse #Observability: from generation log to customer charts #Understand: a warehouse built for agents #Act: continuous evaluation and the Jev Router #Many streams, one store, one loop #Questions this post answers
How much has OpenRouter's request volume grown since it adopted ClickHouse?
OpenRouter's monthly requests grew 33 times, from around 600 million in November 2024 to 19.7 billion by August 2026. The platform now serves over 110 trillion tokens per week, and its raw generation log in ClickHouse holds roughly 134 billion rows. Engineers scaling LLM gateway infrastructure can follow real-world growth benchmarks like this on daily.dev.
What is the Jev Router and how does OpenRouter's model selection work?
The Jev Router, from TypeSafe AI and available since October 2026, lets OpenRouter pick the model itself instead of just the provider. It classifies a prompt into one of about 30 task types, ranks models by real usage data in ClickHouse for that task, filters by user constraints like cost and capability, then chooses an endpoint with fallbacks ready. Developers comparing LLM routing approaches can track announcements like this on daily.dev.
How does OpenRouter let non-technical staff query its ClickHouse data warehouse?
OpenRouter built internal AI agents, accessible mainly through Slack, that read documented dbt models (around 270 of them) and small markdown 'skill' files describing how to query ClickHouse. This lets support staff get answers, like tracing a customer's failing requests end to end, without writing SQL or understanding the underlying data structure. Teams designing agent-driven data access patterns can follow practical examples like this on daily.dev.