Use when writing, reviewing, or auditing code that calls OpenRouter (or a similar multi-provider router like Requesty / Vercel AI Gateway / a LiteLLM proxy) to produce RESEARCH results — evals, LLM-as-judge scoring, agent rollouts, benchmark numbers, or generated/labeled training data. OpenRouter's default routing silently load-balances a "model" across providers that may serve it quantized (down to int4), with different context windows and max-output caps, on different inference engines, and drops unsupported sampling/format params — quietly corrupting results and breaking reproducibility. Triggers: "openrouter", "OPENROUTER_API_KEY", "openrouter.ai/api/v1", provider routing, endpoint pinning, `extra_body={"provider": ...}`, quantization of an API model, "which provider served this", reproducibility of API model outputs, auditing a repo's model-inference hygiene.
2026-07-20