Determined AI vs OpenRouter

OpenRouter has moved well past its origin as a cost-and-routing proxy for LLM completions. The new TTS endpoint routes speech-synthesis models from Mistral, xAI, and Microsoft under a single OpenAI-compatible interface — the same model-agnostic pattern applied for the first time outside text inference. Alongside that, the Presets system (named, versioned sets of models, system prompts, and routing config referenced as @preset/slug) shifts the product from ad-hoc per-call configuration to config-as-code infrastructure. Production customers like Descript are running 13 models in parallel through it, cutting new-model evaluation time from over a week to 1-2 hours.

OpenRouter is becoming a universal AI API router rather than an LLM proxy. The arc is clear: first aggregate LLM models, then add routing primitives (Presets, Fusion), then expand modalities (TTS), then formalize enterprise controls (ZDR, guardrails). Each move raises the switching cost and widens the moat — once a team routes both completions and speech through OpenRouter with shared presets, ripping it out requires reworking every model integration simultaneously. The developer documentation cadence (agent loop guides, evaluation patterns, compliance explainers) is deliberately building the knowledge infrastructure that makes OpenRouter the assumed starting point for multi-model applications.

The next surface expansion is likely image or video generation routing — TTS established the pattern, and the existing catalog already indexes image/video models. An enterprise tier with SLA guarantees around ZDR and uptime is also probable given the compliance documentation effort.