Adding DeepSeek R2 and the new reasoning routing rules
R2 is live across six providers. Here's how our default router decides when reasoning tokens are worth the extra output cost.
Blog
Product notes, routing changes and the occasional benchmark.
Prompt caching is on by default now
Repeated context gets cached at the provider when supported, which is cutting effective input spend 60-80% on the workloads we looked at.
What a year of routing 40T tokens taught us about uptime
Provider outages are more common than their status pages suggest. A look at how failover actually behaves in production.