Following two years of development, US startup TypeSafe AI has released its foundational model, Jev. The company was founded by former OpenAI researcher Diogo Almeida, a co-inventor of RLHF and InstructGPT, alongside Sasha Sheng and Erik Gafni. With Jev, the founding team pursues a fundamentally different direction than conventional generative language models. Instead of producing sentences token by token, the system serves as a dedicated decision model for programmatic pipelines.
Conceptually, Jev is inspired by Daniel Kahneman's concept of fast System 1 thinking. Developers pass unstructured state, such as raw text or JSON objects, along with typed queries to the system. Jev then calculates parallel, probabilistically calibrated answers without producing conversational prose. By treating each query as an immediate logical decision unit, the model aims to simplify execution inside automated control architectures.
The model output is strictly constrained to three core primitives to ensure reliability inside software stacks. These include Choice for selecting from predefined options, and Score for numeric ratings. The third primitive, named Noul, provides calibrated probabilities for binary decisions. Software engineers receive typed outputs that can be integrated directly into applications without fragile regex parsing or excessive output formatting prompts.
By discarding sequential token generation, the architecture achieves substantial latency improvements. Recorded response times range between 70 and 500 milliseconds. Across standard classification benchmarks, Jev operates 20 to 200 times faster than general frontier language models. The absence of streamed token generation removes traditional throughput bottlenecks in multi-step workflows.
The pricing model similarly diverges from industry standards. TypeSafe AI charges 0.042 US dollars per million input tokens, whereas output generation is entirely free of charge. For developers building autonomous agents, operational expenses drop significantly, as routine routing and classification steps inside agent loops become far cheaper. The system effectively positions itself as a fast, deterministic controller for complex agent architectures.