[znnajdla]

My experience mirrors this: services like OpenRouter that promise “failover” are pretty much useless except for sandbox testing because models in production are not really interchangeable. Any production harness doing serious agentic work in production is dependent on more model-specific quirks than you would expect. And even if another model works without errors, performance and efficiency is a whole different story. Even the system prompt can and should be tuned to a model’s preferred speaking style, for example <xml tags> for Claude-like models because they were trained on it, while other models do better with other delimiters. Think of the whole harness, prompt, and model as one system, not really with modular parts that can be swapped out if you care about optimal performance.

[marcyb5st]

I believe part of the LLMOps (I don't like the term, but it is what it is) should be building a failover plan with proper testing that check tools trajectories and such. If you have these then you can sort the good enough models from cheaper to more expensive and have the failover you mentioned.

I saw people bulding a mapping of model->{{prompts}, {tools descriptions}, ...}, but that, to me, it feels extreme. I believe it is the model that needs to adapt to your prompts after a certain point. Models that fail to do so won't get our api requests as they will be out of the failoever roster.

[hamandcheese]

OpenRouter doesn't fail over to a different model, it fails over a different provider of the same model.