Router selection checklist

Use this checklist to evaluate any model router, including Emissary. Not every item will matter for every workload. The point is to know which ones do before you commit. The terms are explained in the router design space.

1. Fit with your workload

• Do you need a router, a gateway, or both? Be clear whether you need a decision about which model, a way to reach and manage models, or both. See routers and gateways.
• Does your traffic vary in difficulty? If almost every request needs the same model, routing won't help much.
• Is the front door open-ended? If your steps are fixed, select models offline instead. See when to route, and when not to.
• Are your workloads agentic or conversational? If so, state and cache matter (see 3).

2. Decision quality

• Is the router performance-based or complexity-based? Does it estimate which models can do the task, or guess how hard the task is from the input?
• Is it model-aware? Was it trained on how the models in your pool actually perform, or does it only map query text to a label?
• How is quality protected? Is there a confidence threshold you control, so cost savings can't override quality?
• Can you see why it routed a request? You need that to trust and tune it.
• How does it adapt when models change? Adding or updating a model shouldn't mean starting over, and ideally it tells you which new models fit your traffic.

3. State and cost

• Is it stateful? Does it know what each model has cached, and use it?
• Does it route across effort as well as models? Changing effort on the same model keeps the cache.
• Is it intelligence-density aware? Does it price in expected output tokens, not just per-token price?
• Does it support custom pricing? Contract rates and discounts should be reflected in the ranking.

4. Modality

• Is it modality-aware? It must not route images to text-only models.
• Do you need it to be multimodal? Only if image content should influence the decision.

5. Integration and operations

• Does it return a recommendation, or execute the call? Which fits your stack?
• Does it work with your SDKs and agents? For example, OpenAI- or Anthropic-compatible APIs, and coding agents.
• Whose provider keys does it use? Yours, or credits you buy from the router.
• What does the routing step cost in latency and money? Measure on your own workload.
• What happens when the router is unavailable? Is there a safe default?

6. Evidence

• Run a trial on your own traffic. See run a trial.
• Treat vendor savings figures as claims, not predictions, until you've measured them.

Ready to try Emissary? Start with get started, and see how it compares in Emissary vs Router.com, vs LiteLLM, vs Not Diamond, and vs OpenRouter.