Cheapest Isn’t Best: How We Rebuilt Provider Routing
What if the cheapest AI inference provider isn't actually the best deal?
For a long time, inference routing has been driven by one dominant metric: price. But cheaper doesn't always mean better, especially when latency, reliability, and output quality come into play.
At OpenRouter, we saw this firsthand. Our previous routing system placed too much weight on advertised prices, often favoring cheaper but slower or less reliable providers over those delivering better overall performance.
So we rebuilt it.
Join Quinn Cembrinski, Strategic Analytics, and Scott Morse, Provider Ecosystem Engineer, from OpenRouter for a behind-the-scenes look at how we redesigned provider routing to optimize for what actually matters: the real cost of a good answer.
We'll dive into:
Why we rebuilt routing: How price-driven routing created the wrong incentives for inference providers.
What actually makes inference efficient: Why latency, throughput, caching, errors, retries, and output quality matter beyond sticker price.
Testing against real traffic: How we compared the new router against the old one and what the data revealed.
What happened after launch: How providers are responding when there's no longer a single metric to optimize for.
What's next: What we've learned, what we've changed since launch, and where provider routing goes from here.
Join us for a technical deep dive into the decisions, trade-offs, and real-world lessons behind OpenRouter's new approach to inference routing.
***
OpenRouter is the unified interface for LLMs: 600+ models from 80+ providers, processing 700T+ monthly. Better prices, better uptime, no subscriptions.