TypeSafe AI launches Jev, a fast structured decision model

TypeSafe AI has released its first System One Model, a new class of frontier models designed to make fast, structured decisions that software can use directly. According to the company's announcement on its blog, the initial model, named Jev, is available today in early access. The core claim is that Jev achieves similar levels of intelligence on System One tasks compared to existing large language models (LLMs) while being two orders of magnitude faster and more efficient. The company positions Jev as a "frontier-intelligence function call" that takes unstructured state in and outputs typed probabilistic decisions, explicitly giving up string generation to optimize for structured outputs and eliminate hallucinations.

The details

The technical differentiator of Jev lies in its architecture and training methodology. TypeSafe AI states that Jev uses a new model architecture, a parallel sampler for maximum efficiency, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD). Unlike existing LLMs that rely on Reinforcement Learning with Human Feedback (RLHF) or Reinforcement Learning with Verifiable Rewards (RLVR), RLCD optimizes for calibrated decisions with epistemically honest probabilities.

Key specifications provided in the launch materials include:

• Speed: End-to-end response times range from 70ms to 500ms, compared to 3 to 329 seconds for frontier LLMs. This represents a 40x to 200x speedup for System One shaped queries.
• Cost: Input tokens are priced at $0.042 per million tokens ($42 per billion tokens). Output tokens are listed as "FREE (too cheap to meter)."
• Output Format: Type-safe structured values with pre-defined possible outputs and structure. The company claims it is mathematically impossible for the model to make type errors.
• Confidence: Every output is accompanied by calibrated probabilities and confidence scores, where higher confidence correlates with higher accuracy.

The company contrasts this with existing LLMs, which generate strings sequentially, one token at a time. Jev generates all outputs in a single query using parallel sampling. TypeSafe AI notes that while strings are flexible and general, they are costly and require parsing and validation when used in software, with a risk of the AI "going off the rails."

Context

Diogo Almeida, the founder of TypeSafe AI, previously worked at OpenAI where he helped build methods for language models to follow instructions, contributing to the research behind ChatGPT. In his manifesto, Almeida states that after four years of asking "where is all the automation?" despite models being superhuman at chat, he concluded that something big was missing. He spent two years in stealth developing this new stack focused entirely on automation.

The launch positions TypeSafe AI against the current trend of general-purpose LLMs that are powerful but require human oversight due to their freedom and potential for hallucination. By constraining the model's freedom through type-safe outputs, TypeSafe AI aims to make these models easier to compose into reliable systems for AI-powered workflows, smart if-statements, and real-time applications where UX is critical.

What it means for developers

For developers, Jev represents a shift from using LLMs as general-purpose text generators to using them as specialized decision functions within existing codebases. The primary use cases cited are:

1. AI-Powered Workflows: Using structured outputs as fuzzy decision rules to classify, route, score, extract, or branch where hand-written logic is too brittle.
2. Real-time Applications: The 70ms-500ms latency allows for AI integration in applications where user experience is critical, such as real-time scoring or guardrails.
3. Verification and Guardrails: Using Jev to score, judge, verify, and detect jailbreaks of LLM prompts, reasoning traces, and outputs.

Developers should note that Jev is optimized for System One tasks—fast, structured decisions—and may not be suitable for tasks requiring open-ended text generation or complex reasoning that benefits from chain-of-thought. The company provides a simplified query example in their blog post, showing how the model handles a dense paragraph of state to produce structured probabilities. They emphasize that the most reliable real-world workflows have many independent, decomposed questions with fine-grained behavior dependent on probabilities.

What to watch

• Sustainability of Pricing: TypeSafe AI acknowledges they cannot prove their pricing isn't subsidized and will need the long-term to prove sustainability, though they expect prices to go down.
• Real-world Performance: While the company claims Jev owns the Pareto frontier for almost 2 orders of magnitude in their new evaluation, developers should test it against their specific production workloads, which may be more complex than the published demos.
• Comparison to GPT-5.6 Terra: The company used GPT-5.6 Terra with default reasoning as a comparable benchmark, noting only one disagreement on "Churn likelihood level" in a side-by-side demo. Independent verification of these benchmarks is necessary.
• Adoption in Automation Pipelines: Watch for third-party integrations and case studies that demonstrate Jev's effectiveness in large-scale, real-time automation scenarios beyond the company's own examples.