Before OpenRouter, picking a model was a wedding. Pick your vendor, sign the terms, bet your product's future on them.
OpenRouter made it a dating app. One API key, one schema, access to essentially every frontier and open model. When Claude is down or slow, traffic fails over to GPT or DeepSeek automatically. When a new model drops, it's in the list the same day.
The underrated features: automatic price discovery (it shows you the actual cost per 1K tokens for every provider version), spend caps, per-key budgets, and unified rate limits.
What we use it for: model comparison during development, failover in production for non-critical paths, and shadow testing new models without rewriting the app.
What we don't use it for: production hot paths where latency matters more than resilience — the extra hop adds ~150ms. For those we pin a provider directly.
The strategic value: you never get locked into one model again. That optionality is worth more than the 10% markup.