OpenRouter launches Jev Router, claiming 82% more completed tasks than its own Auto Router
AI gateway OpenRouter on September 25 launched a model router called Jev Router, which uses TypeSafe AI's decision model Jev to assess how difficult each individual request is — and then send easy tasks to cheap small models and complex tasks to expensive frontier models. The launch comes just ten days after Jev itself emerged from stealth mode on September 15, 2026, and is, according to the coverage, the first major platform integration of the model.
The company behind the model, TypeSafe AI, is simultaneously under an intense spotlight: According to Bloomberg, the launch video for Jev has been viewed roughly 40 million times on X, and Bloomberg cites the Financial Times as saying investors have approached the company with offers valuing it at more than $10 billion. Neither the valuation nor the router's benchmark figures are independently confirmed — worth keeping in mind throughout this story. The coverage of the launch rests on secondary sources that refer to the company's own statements.
How the router works
Jev is not a generative language model. It does not produce text, but returns three types of structured output: "Choice" for selecting one option among candidates, "Score" for evaluating something against criteria, and "Noul" for returning a probability value between 0 and 1. Each response comes with a calibrated confidence score (BigGo Finance).
In Jev Router, this is used to score each request's difficulty in real time and then route the task onward: easy questions go to cheap small models, while heavy tasks are sent to expensive frontier models. Whether this automatic distribution actually reduces costs or improves quality in practice is not documented in the available coverage — that is an assumption, not a proven result.
According to Cryptobriefing, Jev is trained exclusively on synthetic data, using a method the company calls Reinforcement Learning for Calibrated Decisions, abbreviated RLCD (Cryptobriefing). Development is led by Diogo Almeida, who spent nearly four years at OpenAI and is a co-author on the papers behind GPT-4, ChatGPT, and RLHF/InstructGPT (BigGo Finance).
The numbers — with caveats
OpenRouter reports that in benchmark testing where Jev Router and the company's own Auto Router each route four categories of AI agent benchmark tasks, Jev Router completed 82 percent more tasks. The methodology behind the measurement has not been published, and the benchmark most likely comes from OpenRouter itself. It should therefore be read as a company statement, not independent verification (BigGo Finance).
TypeSafe's own numbers are more extreme: The company claims that Jev performs the same task roughly 193 times faster than frontier LLMs, at around 445 times lower cost. These headline figures — 193.6x faster and 444.6x cheaper — come from the company's own workflow evaluations, and the launch post itself acknowledges that they sit at the high end of realistic gains, with GPT-6 Astra and Fable 5.1 as reference (MarkTechPost).
The pricing, by contrast, is concrete: Jev 1.13 costs $0.042 per million input tokens, with no separate charge for output (BigGo Finance).
Earlier benchmark testing at Vercel showed, according to Cryptobriefing, that Jev was 5–18 times faster than certain OpenAI models on classification tasks, and 10–20 times cheaper than Gemini for email classification. Interest was large enough that the API was overloaded at the launch on September 15 (Cryptobriefing).
Privacy
OpenRouter states that Jev only reads the conversation text in order to choose the model and inference strength, and that the entire process follows a zero-data-retention policy: Prompts are not stored or used for training, attachments are never sent to Jev, and requests flagged with "zdr: true" are compatible with the router (BigGo Finance). This is presented as the company's own statement — there is so far no independent review of how the policy is enforced in practice.
An ecosystem in its early phase
Model routing is only one of several early use cases for Jev. LangChain offers this as ModelRouterMiddleware, and there is a jev-router integration that performs the evaluation per turn for Claude Code and Codex (MarkTechPost). OpenRouter is, in other words, one adopter among several — but according to the coverage the first to productize the use case at platform level.
The capital rumors
TypeSafe AI came out of stealth on September 15, 2026, with a $40 million seed round led by DCVC, at a post-money valuation of roughly $200 million. According to Cryptobriefing, the company is now in talks to raise more than $1 billion at a valuation of over $10 billion — an implicit 50-fold increase in under three weeks (Cryptobriefing). Neither the amounts nor the participants have been confirmed by the company; Almeida has reportedly only confirmed significant investor interest. Bloomberg's September 25 story covers the virality and cites the Financial Times on the valuation, but mentions no completed deals (Bloomberg).
The open question
The most persistent uncertainty concerns the quality–cost tradeoff in practice. Cryptobriefing notes that early tests show Jev does not always win on raw accuracy. That means the gain from cheap routing — measured by OpenRouter as 82 percent more completed tasks — depends on whether the difficulty scores actually hit the mark in everyday scenarios, and not just in OpenRouter's own four benchmark categories.
For developers evaluating cheap routing this week, there is now a concrete, usable option. But both the 82 percent figure, the 193x numbers, and the valuation rumors currently rest on the companies' own statements. Independent verification of the benchmarks is lacking, and it remains an open question whether Jev Router's gains hold outside OpenRouter's own test categories.