DeepSeek R1 in Production: A Ship Checklist

Why R1 is different

R1 is built for reasoning-heavy workloads. That usually means longer completions and higher token throughput per successful task than chat-optimized models.

Before you promote to 100% traffic

• Define **success metrics** per surface: cost per resolved ticket, accuracy on a fixed eval set, or human review pass rate.
• Add **hard max_tokens** per route so runaway chains cannot drain balance overnight.
• Log **request ids** end-to-end so you can correlate gateway latency with model latency.

Operating tips

Stage rollouts with a shadow percentage, then shift mix only when P95 latency and error rates stay inside SLO for a full business week.