Diogo Almeida, former OpenAI researcher and co-inventor of ChatGPT, founded TypeSafe AI and launched the new model Jev—a model that defies mainstream trends by not generating text, but instead outputting deterministic structured decisions. Achieving end-to-end latency of 70–500 milliseconds through parallel computing, Jev is 40–200 times faster than existing large models, with an input cost of just $0.042 per million tokens. Positioned as a “System One” model, Jev aims to provide AI agents with a high-speed, low-cost decision layer to address execution friction in enterprise workflows. However, the technology faces skepticism over lack of interpretability and undisclosed technical details, and remains in limited beta testing.

Author and source: Hualin Wuwang

The person most knowledgeable in the AI community about how to make models "talk" has now decided to strip large models of their language abilities.

This former OpenAI researcher and co-inventor of ChatGPT, Diogo Almeida, has returned to the table after two years in stealth, launching his new company TypeSafe AI with $40 million in funding. Their newly released model, Jev, cannot write weekly work reports or engage in late-night chats with users. It can’t even produce a single punctuation mark.

This is a counterintuitive moment. Over the past three years, the entire industry has been teaching large models how to think like humans—slowly. Labs have been desperately extending reasoning chains, prompting models to silently mutter thousands of words before arriving at an answer.

But Jev took a completely different path: it does not generate text, only outputs deterministic structured decisions.

This is not just the launch of a new product—it’s a direct challenge to the current large model technology paradigm.

Give AI a "reflexive response"

To understand what Jev really is, you first need to grasp how difficult current AI developers find their work.

Today, most engineers do something absurd when integrating large models into business systems. You only need it to answer, “Is this email spam?” or “Choose one out of five API endpoints,” yet you’re forced to make a massive model with hundreds of billions of parameters painstakingly type out a string of JSON characters, word by word, over several long seconds.

To ensure the string adheres to code standards, developers must write two full pages of prompts, begging the model to "output no fluff, only pure JSON format." Even then, the model occasionally adds a phrase like "Sure, here is the format you requested" at the beginning or end, instantly breaking the entire downstream pipeline.

Jev's approach to solving the problem is extremely crude. It is not a text autoregressive model at all.

In Kahneman’s renowned book "Thinking, Fast and Slow," human thought is divided into two systems: System One, which is unconscious, rapid, and instinctive; and System Two, which is slow and requires deliberate mental effort. Current mainstream large models are fiercely competing in enhancing System Two, while Jev defines itself as the world’s first pure “System One Model.”

It does not generate content word by word; all decisions are produced simultaneously in a single forward parallel computation. This means it has no generative fluff, and the output is always strictly confined to the schema predefined by the developer.

On a mathematical level, Jev completely eliminated the formatting illusion of structured output.

Since there's no need to sequentially predict the next token, its end-to-end latency is compressed to between 70 and 500 milliseconds, making it 40 to 200 times faster than today’s leading large models on the market.

TypeSafe even had Jev play the shooting game Doom. In scenarios requiring no complex long-term planning, only about 10 instantaneous decisions per second, Jev performed like a highly responsive human player, with the entire reasoning cost totaling approximately $7 per hour.

Even more striking is the pricing: Jev’s input cost is just $0.042 per million tokens, with output completely free. On average, the cost of a single structured decision is approximately $0.0004. For comparison, calling GPT-5.6 Terra costs about 76 times more, while Claude Opus 5 costs hundreds of times more.

Input unstructured data, instantly output classified choices with calibrated probabilities, numerical scores, or boolean judgments, then immediately relinquish control. This is Jev’s entire job.

The agent's bottleneck is not that its brain is too slow

Why did TypeSafe choose to develop a completely non-verbal model at this point in time?

The answer lies in the AI agents being rolled out comprehensively.

Whether it’s Cursor, GitHub Copilot, or enterprise automation workflows from Silicon Valley, engineers quickly realized a harsh reality: today’s agents are often slow and expensive not because their core logic isn’t smart enough, but because of excessive execution friction in between.

A complete agent task typically involves dozens of small decisions, such as determining whether the current step was successful, deciding whether to call Tool A or Tool B, extracting key fields from the previous output, and deciding whether to terminate the process.

If every tiny action requires calling a massive model with hundreds of billions of parameters, latency will accumulate layer by layer to tens of seconds or even minutes, and costs will quickly spiral out of control. Even more critically, if the JSON parsing fails at any single step, the entire Agent will immediately freeze mid-execution.

The industry has been held captive by autoregressive text models for too long, leading most to assume that all AI tasks must go through natural language as an intermediary.

But in real-world production systems, the vast majority of nodes are essentially traditional code control flows. Code doesn't require emotional value, code doesn't need parallelism—code only needs an extremely cheap, extremely fast, and absolutely type-consistent decision signal.

Jev's introduction essentially adds a "pre-reflex arc" layer into the large model's system architecture.

In an ideal collaborative architecture, expensive, cutting-edge large models should step back into the background, serving as System Two to handle high-level planning and complex reasoning. Meanwhile, frontline tasks—such as ticket classification, initial content compliance screening, bulk data labeling, and precisely selecting which of dozens of APIs to call—can be effortlessly completed in milliseconds by System One models like Jev.

TypeSafe's internal test data shows that Jev achieved an accuracy rate of 67.8% across four typical enterprise workflows, nearly matching GPT-5.6 Terra's 67.9%. When considering only judgment precision, it does not lag behind large models that cost tens of times more, while its execution speed is dozens of times faster.

Separating the slow, deliberate thinking of the cerebral cortex from the reflexive responses of the spinal cord may be the right approach for AI agents to truly achieve large-scale industrial deployment.

The black box behind zero error

However, it is premature to view Jev as a perfect silver bullet. Behind this impressive technical achievement, at least two significant shadows have been deliberately avoided.

First is the complete loss of interpretability.

Although traditional chain-of-thought models are verbose and prone to errors, they at least document the reasoning behind "why this choice was made" on scratch paper. When a credit approval agent rejects a customer’s application or a security system blocks a transaction, the compliance team can trace the model’s decision-making logic through this reasoning process.

Jev’s output consists only of dry options and probabilities; it lacks the ability to explain its reasoning in natural language. When the model makes a critical misjudgment, no one can trace the cause from its black-box weights. In highly regulated fields such as finance, healthcare, and law, this lack of auditability often leads directly to compliance failures.

Secondly, the technical details and evaluation benchmarks are closed.

TypeSafe claims to have invented a novel training method called RLCD (Reinforcement Learning based on Calibrated Decisions), replacing traditional RLHF. However, based on currently available information, the official team has not disclosed details such as how the reward function is designed, what underlying network architecture is used, or the mathematical methodology behind calibration probabilities.

More subtly, consider those impressive scores: all current benchmarks originate from internal evaluations by TypeSafe, with no independent third-party replication completed. Even the reference answers used as ground truth were generated by averaging outputs from GPT-6 Astra and Claude Fable 5.1. Using competitors’ next-generation models as judges, then declaring overwhelming victory in cost-performance, creates a self-reinforcing loop that carries a strong public relations flavor.

Not to mention the sustainability of the business model. Charging just four cents per million tokens with completely free outputs is so low that it raises suspicions of burning through $40 million in funding to wage a price war.

Jev is still in a limited invite-only beta phase; it remains uncertain whether this system can maintain both ultra-low latency and financial balance when faced with large volumes of concurrent enterprise-level traffic.

Making models smaller, faster, and more deterministic does address current industry pain points. However, it's still too early to claim that it has fundamentally disrupted the paradigm before it has been truly stress-tested under extreme real-world business conditions.

Over the past few years, everyone has been amazed at how increasingly articulate AI has become. Now, finally, some are realizing that in the silent world of machines, actions often matter more than words.