TypeSafe AI

What is TypeSafe AI?

TypeSafe AI is a San Francisco-based frontier AI lab focused on building intelligence for machines rather than humans. TypeSafe AI is an AI lab building machine-native intelligence infrastructure for automation, designed to make decisions within software. Rather than another conversational chatbot, the company is pursuing what it calls “System One Models” — small, fast models meant to be embedded directly inside applications, pipelines and agents to make split-second judgments that regular code can act on.

The company’s flagship release is Jev, its first public System One Model. System One Models are a new class of AI model built for decisions inside software, and Jev is TypeSafe’s first public System One Model, optimized for automation, so developers can send it structured questions and get typed decisions with probabilities and confidence that their software can act on. Unlike a general-purpose LLM, Jev can answer yes/no questions with probabilities, choose among defined options or score an input against a scale, and for the choice and scoring formats it also returns a confidence measure alongside the probabilities. TypeSafe argues this constrained output format eliminates a major source of unpredictability in AI-driven automation, since TypeSafe says Jev “can’t hallucinate” because it only selects from allowed answers.

Performance and cost are central to the pitch. TypeSafe prices Jev at $42 per billion input tokens, or $0.042 per million, with output free because the architecture makes it too inexpensive to meter, and reports end-to-end response times between 70 and 500 milliseconds, with gains of 40x to 200x against conventional models on tasks shaped for Jev. Independent coverage has noted these figures come from TypeSafe’s own internal workflow evaluations covering scenarios like security incidents, agent-trace monitoring, invoice processing and customer service, evaluations that are internal. Confidence scores are a key part of the design: confidence scores could give developers a way to control how much they automate, with TypeSafe recommending automatic action for clear cases, confirmation or more information when confidence drops, and escalation for uncertain cases.

The company was founded in 2024 and headquartered in San Francisco, and has raised approximately $40 million in funding led by DCVC. Its leadership team includes CEO Diogo Almeida, CTO Erik Gafni and COO Sasha Sheng, with Almeida having previously helped develop reinforcement learning with human feedback at OpenAI, Gafni bringing experience building production AI systems, and Sheng having worked in research engineering at Meta/FAIR. Jev is currently in a gated rollout: the model is currently available in early access for select developers, can process hundreds of outputs in parallel from a single prompt, and provides calibrated confidence scores allowing developers to determine when software can act autonomously and when it should escalate.

TypeSafe is squarely aimed at engineering teams that want to bolt semantic judgment onto deterministic systems rather than build another human-facing assistant. As reporting has noted, TypeSafe is pitching Jev for quick judgments such as sorting requests, scoring records and screening AI outputs for jailbreak attempts. Because the tool is so new and still gated behind a waitlist, real-world performance under production load, long-term pricing stability and accuracy benchmarks from independent third parties are not yet broadly available.

How you can use it

Because Jev returns constrained, typed decisions instead of free text, it is designed for scenarios where software needs a fast semantic judgment rather than a conversational answer.

• Security and compliance triage: quickly scoring or classifying security incidents and compliance events for downstream automated handling.
• Agent oversight: reviewing or monitoring the outputs and traces of other AI agents before accepting their results, including screening for jailbreak attempts.
• Back-office automation: processing invoices or support tickets by sorting, scoring, or routing them based on confidence-weighted decisions.
• Customer service routing: making rapid yes/no or multiple-choice decisions to direct customer requests without invoking a full conversational model.

Plans & pricing

Jev is priced on a usage basis rather than via traditional subscription tiers. TypeSafe prices Jev at $42 per billion input tokens, or $0.042 per million, and says output is free because the architecture makes it too inexpensive to meter.

• At Jev's advertised rate, an app making one million calls with an average of 1,000 billable input tokens per call would spend $42, before infrastructure and review costs.
• Access is currently gated: the model is currently available in early access for select developers.
• No public self-serve free tier or subscription pricing page has been confirmed at this time.