Jev Router Selects AI Models By Task Difficulty On OpenRouter
Routing has been priced on task type alone, so a prompt that needed a large model and one that did not cost the same; Jev Router prices the decision itself and reports 82% more benchmark tasks completed than OpenRouter's Auto Router.
Key Facts
- Jev Router became available on OpenRouter on September 28, 2026.
- Jev Router completed 82% more tasks than OpenRouter's Auto Router across four AI agent benchmarks.
- Jev was released by TypeSafe AI on September 15, 2026, and returns decisions with confidence scores.
- TypeSafe AI prices Jev at $0.042 per million input tokens, with output free.
- TypeSafe AI is led by CEO Diogo Almeida, a co-developer of ChatGPT.
Reporting from 2 sources: ASCII.jp, GIGAZINE.
TypeSafe AI's Jev Router became available on OpenRouter, the AI gateway service, on September 28, 2026. The router sends a user's prompt through Jev, the judgment model TypeSafe AI released on September 15, 2026, to score the difficulty and accuracy of the task, then assigns an appropriate AI model. Easy tasks go to inexpensive small models and difficult tasks to more expensive state-of-the-art models, cutting the waste that comes from routing easy and difficult work to the same model. OpenRouter's earlier routers analyze the type of task but not its difficulty. Testing across four AI agent benchmarks found Jev Router completed 82% more tasks than OpenRouter's Auto Router. Jev Router does not save prompts and does not use them for training. TypeSafe AI is led by CEO Diogo Almeida, who contributed to the development of ChatGPT.
- Jev is described as a "System One model," named for the fast, intuitive judgment Nobel Prize-winning economist Daniel Kahneman called System 1, distinct from slower System 2 deliberation.
- TypeSafe AI describes Jev as "a new class of state-of-the-art model built to make rapid, structured decisions that software can use directly."
- Jev uses a training method TypeSafe AI calls "Reinforcement Learning from Correct Decisions (RLCD)," and returns only the decision rather than text.
- ASCII.jp reports Jev costs $0.042 per million input tokens, with output free. A roughly two-minute test battle using about 3 million tokens cost $0.126, about 19 yen.
- New accounts get $5 worth of tokens free, which ASCII.jp's tester had not used up after extensive use.
- HTML implementing Jev for game NPC control and debugging, including a battle where the player character sent two kinds of input every 0.2 seconds.
- GIGAZINE links Jev to earlier coverage describing it as capable of "GPT-5.6 Terra-class performance" cheaply and at ultra-high speed.
- A separate compact decision model, "Jeff," runs locally at about 22 milliseconds per decision on an RTX PRO 6000 and about 28 milliseconds on an Apple M4 Max.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.
Sources
- ASCII.jp AIが人間のようにゲームをプレイ 話題の「Jev」を試してわかったこと
- GIGAZINE ベンチマークでJev超えな意思決定モデル「d1」が登場