Jev: The Anti-LLM Has Arrived

Jev, a new AI from OpenAI veteran Diogo Almeida, offers a fundamentally different approach to machine intelligence. Almeida, whose research built ChatGPT, now leads TypeSafe AI, bringing us what they call a 'System One' model. Unlike 'System Two' LLMs, which mimic slow, deliberate reasoning, Jev makes fast, intuitive decisions.

This isn't another text generator. Jev functions as a specialized classifier, ingesting unstructured data—like an email—and outputting a structured, probabilistic decision. It delivers type-safe outputs, never generating a single word of text. This non-generative architecture completely eliminates hallucinations, a persistent problem with traditional LLMs and their unpredictable outputs.

The practical benefits are stark. Jev is up to 400x cheaper and 200x faster than LLMs for classification tasks, processing millions of input tokens for just $0.042. Its response latencies of 70-500 milliseconds drastically outperform conversational models, which can take 3-329 seconds, while its non-generative architecture completely eliminates hallucinations.

Jev isn't for chatting. Developers provide an input, often a JSON object containing raw data, and define a precise output schema. Jev then processes this information, returning a perfectly structured JSON output that matches the schema exactly, complete with calibrated confidence scores for each decision it makes. This predictable structure is a core benefit.

Three core decision types drive Jev's decision-making. Choice allows it to select one option from a predefined list, such as categorizing an email as "shopping," "work," or "marketing." Score lets it rate an input on a specified numerical scale, for instance, assigning a priority from 1 to 5 or a spam likelihood percentage. Finally, Noul provides a probabilistic true/false answer, like "is this user likely to buy?" Each decision includes a probability, giving developers fine-grained control to set thresholds for automation or human intervention.

Crucially, Jev delivers type-safe outputs. This means its structured JSON responses can be integrated directly into applications without requiring additional parsing, cleaning, or validation. This eliminates common integration headaches, dramatically simplifies developer workflows, and significantly boosts the reliability of systems built using Jev. No more guessing what the AI will return.

Jev truly shines where speed and cost are king, making it indispensable for high-volume, repetitive decisions. Imagine instantly sifting through millions of emails, routing customer support tickets in real-time, or moderating vast amounts of user-generated content. It can even score sales leads with incredible efficiency, a task that traditionally bogs down expensive LLMs. Ryan Vogel, from OpenCode, demonstrated its power by classifying 1,700 emails for just 18 cents, highlighting its unmatched value.

Beyond simple classification, Jev becomes the essential connective tissue for sophisticated AI agents. Use it to validate tool calls, ensuring your agent only executes safe and relevant actions, or screen user prompts for potential jailbreaks, adding a critical security layer. Crucially, Jev acts as an intelligent, low-cost router, deciding if a task genuinely needs an expensive, slow LLM, or if a quick, precise Jev decision will suffice, saving significant operational costs.

Think of Jev as the smart if-statement for your entire application. It embeds nuanced AI judgment directly into your programmatic logic, operating where traditional code simply can't handle the complexity of ambiguous inputs. This allows developers to build more responsive, reliable, and intelligent systems without the overhead of larger models. For more technical details on its architecture and capabilities, check out TypeSafe AI, the company founded by OpenAI veteran Diogo Almeida.

Jev's arrival doesn't mean ditching your LLMs. Instead, it heralds a crucial shift: abandoning the idea that one giant AI model can do everything well. Jev, a System One Model from Diogo Almeida, excels at rapid, intuitive decisions, offering 40-200x faster processing and 400x lower costs for classification tasks. Larger LLMs remain indispensable for complex, deliberative reasoning and creative generation.

This signals a broader architectural trend toward multi-agent AI systems. Picture small, specialized models like Jev handling the vast majority of routine classifications and decisions—from moderating user content to scoring sales leads. Only when a decision requires deeper understanding or open-ended generation does the system escalate to a more resource-intensive LLM. This "smart if-statement" approach optimizes both performance and budget.

For builders, this unlocks unprecedented opportunities to create robust, scalable, and significantly more cost-effective AI applications. The winning strategy isn't about identifying the single "best" model anymore. It’s about intelligently orchestrating a network of specialized AIs, each performing its task with peak efficiency, precision, and a clear confidence score. This new paradigm promises real-world value and a powerful competitive edge.

What is Jev?

Jev is a new type of AI model called a 'System One Model' created by TypeSafe AI. It's designed for fast, structured decision-making, taking unstructured data as input and producing probabilistic, type-safe outputs, rather than generating text like an LLM.

How is Jev different from an LLM like ChatGPT?

Jev is fundamentally different. LLMs generate text in a slow, token-by-token process ('System Two' thinking). Jev makes decisions in a single, parallel query ('System One' thinking), outputting structured data. It is significantly faster, cheaper, and cannot 'hallucinate' because it doesn't generate free-form text.

What are the main use cases for Jev?

Jev excels at high-volume, repetitive classification and decision tasks. Key use cases include routing support tickets, moderating content, classifying documents, scoring leads, and validating tool calls or screening for jailbreaks within AI agent workflows.

Who created Jev?

Jev was created by Diogo Almeida, an OpenAI veteran and co-inventor of RLHF (Reinforcement Learning from Human Feedback), a foundational technology behind ChatGPT. His company is called TypeSafe AI.