TypeSafe AI Launches Jev, a Low-Cost AI Model Built for Software Automation

TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, has introduced Jev, a new AI model designed to make software automation faster and less expensive. Unlike conventional large language models, Jev produces probability-based decisions rather than generating text.

Almeida, who helped develop ChatGPT and reinforcement learning from human feedback (RLHF), founded TypeSafe AI after concluding that language-focused models were not always suited to software automation. The company’s new transformer-based model is designed to provide predefined outputs with confidence scores, allowing developers to use its decisions directly in automated workflows.

How Jev Differs From Traditional AI Models

Jev uses what TypeSafe calls calibrated decisions, returning probabilities instead of written responses. Because developers define the available outputs in advance, the model cannot generate arbitrary text, although its predictions can still be incorrect.

The company says Jev offers free output tokens and charges for input tokens by the billion, not the million. Its launch attracted enough developer interest to temporarily disrupt API access as demand exceeded the company’s serving capacity.

Early users have reported faster, cheaper performance for specific classification tasks. Vercel software engineer Pranit Sharma reported that replacing an OpenAI model with Jev for command safety classification produced results five to 18 times faster, with improved accuracy in the company’s test.

Bryo AI CTO Nikhil Mudholkar also compared Jev with Google’s Gemini for business email classification. Gemini was slightly more accurate in his test but cost 10 to 20 times more, while Jev’s probability scores offered a useful way to determine whether automated workflows should proceed.

AI Agent Monitoring and Model Routing

TypeSafe sees applications for Jev beyond classification, including monitoring AI agents for unsafe behaviour and potential jailbreak attempts. Its low operating cost could make it practical to check agent activity without requiring another large language model to review every action.

Armin Ronacher, CTO of Earendil, which develops the open-source AI agent harness Pi, noted that developers must still decide how to interpret Jev’s confidence scores. A low-confidence prediction may require additional review, while a higher probability could support an automated decision.

Ronacher also identified AI model routing as a potential application. Jev could help determine which model should handle a particular task without the expense of using a larger language model to make that selection.

TypeSafe AI’s Approach to Model Training

TypeSafe describes Jev as a System One model focused on fast, intuitive decisions rather than extended reasoning. Almeida says the company trains the model exclusively on synthetic data using a method called reinforcement learning from calibrated decisions, an approach discussed in its research on specialised AI models.

The company has not disclosed Jev’s full architecture. Almeida says TypeSafe plans to develop additional versions of the model for other modalities, extending its approach beyond the classification and automation tasks supported by the current release.