NVIDIA NeMo vs OpenRouter

OpenRouter has moved from a pure model-routing layer to a broader developer platform. September's changelog covers Ori Eval (an LLM evaluation system with LLM-as-judge scoring), an Image API spanning 20 generation models, a TTS endpoint aggregating Mistral, xAI, and Microsoft, compound model synthesis (Fusion), config-as-code Presets, and Zero Data Retention at the API level. The changelog reads almost entirely as tutorials and case studies — the surface area has grown faster than the documentation and they're backfilling developer enablement.

OpenRouter is constructing the full AI development stack on top of its routing layer: evaluate, configure, generate, score, and synthesize — all through one API. The Descript case study is the strongest evidence: Ori Eval cutting model evaluation cycles from a week to one or two hours is a real workflow change. If that evaluation infrastructure matures into structured regression testing and cost/quality dashboards, OpenRouter becomes the platform where developers both run and benchmark models.

Ori Eval is the highest-leverage product in the catalog. The next visible move is likely tighter Presets integration — letting evaluation configs version alongside routing configs — and automated regression alerts when a model update shifts quality scores on existing test sets.