OpenRouter Rebuilds Provider Routing for Speed
OpenRouter replaced its mostly price-driven provider router with a real-time system that weighs effective cost, latency, throughput, reliability, and response quality. A 24-million-request experiment improved median TTFT by 16.9% and reduced 429 errors by 46%, while increasing request costs by 6%.
This is a meaningful infrastructure upgrade: OpenRouter is optimizing for the cost of a successful answer, not merely the cheapest attempt. The tradeoff is sensible for production workloads, but developers should monitor spend rather than assume “default routing” still means lowest price.
- –Providers now compete on measurable inference quality, speed, capacity, and reliability—not just token pricing.
- –The router remains cache-aware, keeping ongoing conversations pinned to the same provider when possible.
- –Developers retain control through `provider.sort: "price"`, `:floor`, `provider.order`, and provider allowlists.
- –The 6% cost increase may pay for itself in agent workflows by reducing retries, waiting, and failed requests.
- –The quality signal currently emphasizes usable tool calls, so broader output-quality evaluation will determine how powerful this becomes.
DISCOVERED
2h ago
2026-10-29
PUBLISHED
3h ago
2026-10-29
RELEVANCE
AUTHOR
OpenRouter