TypeSafe's Jev Router picks models for free, with a claimed million-token window

Why it matters

Jev Router turns TypeSafe's decision-model thesis into a model-selection product, but its free price and million-token context differ sharply from the separate Jev 1.13 listing. OpenRouter has not yet published usage data or comparative routing results for the new endpoint.

TypeSafe's Jev Router, listed on OpenRouter on September 25th, says it can select a model and reasoning effort for each request, and the listing prices it at zero. The product is built around Jev, TypeSafe's model for structured decisions, and is framed as one endpoint for accessing a broader model ecosystem.

That makes this a different proposition from simply offering Jev through another API. In its September 15th launch post, TypeSafe founder Diogo Almeida described Jev as a model that returns typed decisions instead of generated text. Almeida previously worked at OpenAI and Google Brain; TypeSafe credits him with co-inventing RLHF and InstructGPT. His stated problem was the gap between language models' strength at conversation and software's need for dependable automated decisions.

The new listing puts Jev in the position of choosing which model should handle a request. OpenRouter describes Jev Router as adapting to the conversation and balancing quality, speed and cost. The page says it runs on Jev and that one endpoint can draw on the wider model ecosystem's token and spend share. That description does not explain which models are eligible, how the router weighs those tradeoffs, or how a developer can constrain its choices. The listing also shows no usage data yet, so there is no public evidence there of which models Jev Router selects in practice or how often developers use it.

OpenRouter already offers its own Auto Router, which selects models using community spending patterns over a trailing seven-day window. Jev Router enters the same general problem from a different direction: TypeSafe presents its decision model as the selection layer, while OpenRouter describes its router as drawing on its marketplace's observed usage. The listing does not publish a head-to-head evaluation, so developers cannot yet compare their routing quality, latency or total cost from the available figures.

The price and context claims need to be read as properties of this router listing, not assumed to describe every Jev request. OpenRouter's separate Jev 1.13 listing gives that model a 32,000-token context and charges $0.042 per million input tokens, with no charge for output. The Jev Router page instead says the router is free and lists a 1,000,000-token context window. It does not explain the difference between those figures or whether the router's advertised window applies to the routed model, the routing process, or both.

The page also lists audio, PDFs and other files, images, text and video as supported inputs, with text as the output. That broad input menu sits alongside a product description centered on choosing models and reasoning effort. The practical value will depend on what happens after that choice: whether the selected model can handle the input, whether the route reduces work for application developers, and whether the choice improves cost or quality over a fixed model selection.

For Almeida, the launch extends TypeSafe's original bet from a decision model to a decision layer for model access. Jev was introduced as an alternative to asking a text-generating model to make bounded choices inside software. Jev Router applies that idea to a higher-level choice: which model should answer. The OpenRouter listing makes that proposition available through its platform, but leaves the selection rules and performance evidence unspecified. For now, the clearest claims are the interface, the zero price shown on the page and its advertised context window; the listing's own usage panel has no data yet.