While generative AI has set off an arms race in text and multimodal models in Silicon Valley, Diogo Almeida, a former core OpenAI researcher who helped build ChatGPT and its reinforcement learning from human feedback (RLHF) architecture, has chosen to lead his startup TypeSafe AI down a completely different technical path. On September 18, 2026, the company officially launched "Jev," a new Transformer model built specifically for "decisions," instantly sparking lively discussion across developer communities.

The breakthrough marks a major shift in thinking about enterprise AI architecture, because the industry has long been plagued by the "hallucinations" and high inference costs of large language models (LLMs). TypeSafe AI boldly abandoned the traditional word-by-word text-generation mechanism and focused instead on pure logical judgment and probability output, aiming to build a nerve center for cloud infrastructure and software engineering that is both fast and completely reliable.

For enterprise CTOs actively pushing digital transformation and building multi-agent systems, Jev's arrival is a real shot in the arm. The model not only uses extremely efficient algorithms to remove the uncertainty traditional models bring to automated workflows, but also upends the market's rules with an unprecedented pricing model, opening new possibilities for information architecture worldwide.

Key takeaways

  • A revolutionary decision architecture: TypeSafe AI's decision model Jev uses a "System One" architecture that outputs only precise probability distributions, eliminating hallucinations by design.
  • Peak performance at ultra-low cost: it runs 5 to 18 times faster than OpenAI Luna 5.6 at a few percent of the cost of top models such as Gemini, and output tokens are completely free.
  • Core use cases: software workflow automation, intelligent routing and guardrails for multi-agent systems, deeply trained with synthetic data and calibrated decision reinforcement learning.

Inside Jev's technology: leaving text generation behind for a pure "System One" architecture

Jev completely overturns the operating paradigm of traditional large language models, turning the Transformer architecture from lengthy text generation to highly precise probabilistic decisions. Traditional LLMs rely on autoregressive algorithms that predict the next word one at a time. That mechanism shines in writing and conversation but easily produces fatal hallucinations when making high-reliability business decisions. Led by Diogo Almeida, Jev drops the baggage of text generation altogether and borrows "System One" from cognitive psychology, focusing on fast, intuitive, zero-latency logical judgments. When a request reaches the server, the model outputs only a predefined, precise probability distribution and confidence score, giving developers fully deterministic structured results in their back-end software.

Combining calibrated decision reinforcement learning with purely synthetic data is the key technical pillar behind Jev's highly reliable decisions. To keep its probability outputs rigorous, TypeSafe AI completely dropped noisy public web text during training and used only highly structured, high-quality synthetic data. Through its proprietary "calibrated decision reinforcement learning," the model learns to strictly evaluate its own judgments and define their boundaries, mathematically eliminating the risk of fabricated facts. With this architecture, engineers watching their monitors no longer have to nervously write complex filtering logic for unexpected strings, greatly improving the stability of the whole system.

Market impact and a cost revolution: free output breaks the pricing myth of top LLMs

TypeSafe AI's disruptive pricing and peak performance are hitting the current cloud AI services market directly. In today's mainstream market, whether handling simple intent recognition or complex text generation, companies pay steep bills for huge numbers of input and output tokens. Jev instead uses a model rarely seen in the industry: output tokens are free, and input tokens are priced per billion. That pushes the marginal cost of inference down to a few percent of top commercial models such as Gemini. With this highly aggressive strategy, large-scale real-time information processing is no longer the preserve of tech giants; small and medium-sized businesses can easily afford massive request volumes.

In measured performance, Jev shows a crushing advantage of several to more than ten times over traditional large models. According to test data from early adopter Vercel, when running high-frequency tasks such as large-scale workflow automation and filtering instructions for safety, Jev responds 5 to 18 times faster than OpenAI Luna 5.6. For modern web applications and API gateways that need millisecond responses, this sharp drop in latency means companies can embed multiple layers of intelligent decision checkpoints across the whole request path without hurting the user experience, completely reshaping architecture standards for high-concurrency systems.

From guardrails to intelligent routing: infrastructure bedrock for the multi-agent era

Jev's core value is serving as the most critical safety guardrail and intelligent model router in complex multi-agent ecosystems. As autonomous agents spread through enterprise automation, intercepting malicious prompt injection in real time and keeping agents on track has become the top problem for security and engineering architects worldwide. With its millisecond intuitive decisions, Jev can assess request risk at the system's entry point and decide instantly whether to block or allow it. It can also act as a traffic splitter, leaving simple queries to lightweight models and calling expensive large models only when necessary, forming a highly precise intelligent routing network.

For Taiwanese and Asian markets that depend heavily on high-tech supply chains and digital commerce, Jev offers a cost-effective path to upgrade automation. When Taiwanese companies push internal digital transformation, they often face the twin challenges of costly computing hardware and strict security compliance. Jev's lightweight, low-latency design can integrate seamlessly into existing ERP, CRM systems and production-line edge devices, helping companies build secure and efficient automated decision pipelines without expensive compute clusters and raising the competitiveness of the whole industry.

Frequently asked questions (FAQ)

How is Jev fundamentally different from traditional large language models (LLMs)?

Jev gives up the autoregressive text generation of traditional LLMs and focuses instead on structured, intuitive decisions. Traditional language models are built around predicting natural language word by word, so they carry the risk of "hallucinating" false content. Jev uses a System One architecture: after receiving input, it outputs only a predefined probability distribution and a decision confidence score. That removes the hallucination problem by design, making it purpose-built for deterministic decisions and logical filtering.

In which specific business scenarios is Jev the best fit?

Jev is best suited to high-concurrency workflow automation, safety guardrails for multi-agent systems and intelligent model routing. Whether filtering malicious prompts in real time, judging user intent to route traffic to different back-end servers, or letting automation software make precise conditional branching decisions in multi-step workflows, Jev can make the call 5 to 18 times faster than OpenAI Luna 5.6 and at very low cost.

How do Jev's unusual pricing and training help companies lower their total cost of ownership?

Jev uses an ultra-low-cost strategy with completely free output tokens and input priced in units of one billion tokens, combined with training on high-quality synthetic data. When companies process huge volumes of real-time information, inference costs are only a few percent of top models such as Gemini, and because training uses calibrated decision reinforcement learning throughout, engineers spend far less time later tuning prompts and debugging.

Conclusion: a new era of AI that moves from generative fever to rational decisions

The birth of TypeSafe AI and Jev marks the AI industry's return from a fever for ever-larger parameter counts and generation ability to a rational engineering era focused on practicality, reliability and maximum efficiency. Led by a former core OpenAI scientist, this decision revolution proves that "giving up generation" can buy unmatched speed and precision, and it lays a solid safety foundation for the next stage of autonomous multi-agent systems. As digital technology keeps evolving, precise models specialized in intuitive decisions are bound to complement general-purpose generative models, together building a sturdier and more commercially sensible blueprint for next-generation intelligent computing.