TypeSafe: Jev Router

TypeSafe: Jev Router (typesafe/jev-router) is a hosted, cache-aware model router on OpenRouter, not a single generative LLM. Before each turn it runs Jev, TypeSafe’s first System One decision model, which reads conversation text (under zero-data-retention terms) and scores task difficulty, precision needs, and whether a stronger model or higher reasoning effort is warranted. It then selects a downstream model and reasoning setting from OpenRouter’s curated pool, preferring the cheapest option that meets the bar while keeping a working model across the session so prompt-cache value is not discarded lightly.

Clients call it like any other OpenRouter chat model via the OpenAI-compatible Chat Completions API. Billing follows whichever backend model serves the reply; the router itself has no fixed per-token price. OpenRouter advertises routing to models with up to 1,000,000 tokens of context, multimodal inputs (text, images, audio, video, and files such as PDFs), tool use, structured outputs, and adjustable reasoning effort. Optional jev-router plugin fields (models, allowed_models, excluded_models) narrow the candidate pool without changing Jev’s ranking logic.

In OpenRouter’s September 2026 launch comparison across four agent benchmarks (423 tasks total), Jev Router solved 237 tasks versus 130 for OpenRouter’s Auto Router—a platform-reported result, not a standard knowledge or coding leaderboard score for one fixed checkpoint. Independent Agent Arena runs on real agent sessions reported modest net gains and strong steerability at higher cost and latency than calling a single frontier model directly. Use Jev Router when you want one endpoint that adapts model choice and reasoning over long, evolving agent or chat workflows rather than a static model card.