Jev Router

OpenRouter launched typesafe/jev-router on September 25, 2026, as a model-routing tool powered by TypeSafe AI's Jev decision model. The router uses Jev to instantly assess each request's difficulty, then routes simple tasks to inexpensive small models and complex tasks to high-priced frontier models, maintaining output quality while reducing API costs. Unlike the existing Auto Router, which only identifies task type, Jev Router analyzes task difficulty variations within the same task category. The router reads conversation text solely for model and inference-strength selection, with a zero-data-retention policy: prompts are not stored or used for training, attachments are never sent to Jev, and requests marked with "zdr: true" remain compatible. Benchmark tests reported in the coverage show Jev Router completed 82% more tasks than OpenRouter's existing Auto Router. TypeSafe AI is the model creator; OpenRouter serves as the distribution host.

OpenRouter released typesafe/jev-router as a free model router that selects the optimal model and reasoning effort per request. Its distinguishing feature is that it "calculates cache losses before switching models," prioritizing retention of warm prompt caches over naive cheapest-model routing. Because each provider caches the stable prefix (system prompt, tool definitions, schemas), switching models mid-session forces expensive reprocessing of repeated tokens. The analysis illustrates the economics with Claude Sonnet 4.6, where a cached read runs roughly $0.30 per million tokens against $3.00 per million for fresh input, a ten-to-one gap. By pricing that cache-retention cost before deciding to switch, Jev Router reframes inference cost as a discount-based rather than list-price calculation. The router is a TypeSafe AI model served through OpenRouter's aggregator.