The past weekend saw a model named Jev rapidly gain traction across the developer community.

It cannot engage in casual conversation like ChatGPT, cannot write long-form articles, and will not dutifully generate code. Instead, its function resembles a high-frequency multiple-choice test: continue or stop? Approve or intercept? Which button to click? Which employee should receive this email?

On the surface, its capabilities seem significantly reduced, so why has it become so popular? Because Jev targets one specific task: making judgments based on a zero-hallucination foundation.

Jev was spearheaded by Diogo Almeida, co-creator of ChatGPT. He wrote on X: "After co-inventing ChatGPT, I kept asking myself: why haven't superhuman-level chat models delivered AGI?" After two years of stealth development, he offers his answer - a novel decision-making model that may prove more effective than even more advanced chatbots. According to benchmark data published on TypeSafe's official site, Jev executes automated workflow tasks 193.6 times faster and at 444.6 times lower cost than existing large language models. Input costs are just $0.042 per million tokens, with free output tokens. Since launch, demand has exceeded expectations, briefly overwhelming the API.

Presently, Jev is accessible to all users without a waitlist, and registration grants $5 in credits (approximately 120 million tokens).

Two Contradictory Paths in One Career

Diogo Almeida's professional background carries an inherent tension. At OpenAI, he helped construct RLHF (Reinforcement Learning from Human Feedback) and InstructGPT, methods that directly led to ChatGPT's creation. Then he departed, beginning to question this very methodology.

In one presentation, he directly questioned: "Today's AI can tackle advanced mathematics, so why are enterprises still hesitant to entrust customer service decisions to it?" His answer: "Today's AI is exceptional at assistance, but not yet proficient at automation."

Assistance means a human sits at the screen while AI does the work and the human verifies it. Automation means no human is present, as AI makes and executes decisions autonomously in the background.

Almeida traced the issue back to RLHF itself. He stated during a talk: "Why do today's large models always require a human in the loop? Because when we train them, we literally embed humans into that loop."

RLHF teaches models one lesson: which responses do humans prefer? This is advantageous for chat products. But for automation systems, it becomes the core problem.

Decoding Jev: A Function Designed for Judgment

TypeSafe's website describes Jev as: "Taking unstructured state input and outputting typed probabilistic decisions." In simpler terms: feed it information, and it tells you the answer and its level of confidence, all without generating any hallucinations.

It does not generate text. It performs three core functions:

Noul: Boolean judgments, returning a probability between 0 and 1 representing the likelihood of a statement being true.

Choice: Selects one option from a list of up to 255 predefined choices, suited for classification and routing.

Score: Grades output on a defined scale for assessing urgency, quality, or risk.

Every response includes a complete probability distribution and confidence metric. It offers strongly typed output, eliminating the need for JSON prompts or additional parsers.

Examining the Speed and Cost Advantages

Standard large language models generate responses sequentially, one token at a time, with each dependent on the previous. Jev's sampling operates in parallel, generating all outputs in a single query. This fundamental architectural difference explains the speed gap.

According to TypeSafe's data: Jev's end-to-end response time is between 70 and 500 milliseconds. Existing frontier models range from 3 seconds up to 329 seconds. This creates a 40x to 200x speed difference on System One type tasks. Price-wise: Jev's input costs $0.042 per million tokens ($42 per billion tokens). Existing frontier models cost between $0.20 and $10 per million tokens for input. Output tokens are free.

TypeSafe notes that the 193.6x speed and 444.6x cost improvements come from their automated workflow benchmarks, describing them as "expected to be on the higher end of real-world benefits." These evaluations were produced by the company's own model capability team, so some bias is possible.

Calibration Is the Real Foundation

Beyond speed and price, TypeSafe emphasizes "calibration." Current large language models often exhibit overconfidence and inconsistency when providing confidence scores. TypeSafe's website states: "If a model can complete a task 95% of the time, but cannot tell you when it is in the other 5%, it cannot automate that task."

To address this, TypeSafe developed RLCD (Reinforcement Learning for Calibrated Decisions). The goal is that if the model states an 80% probability for an event, then across all events it judges at 80%, roughly 80% should ultimately occur. This directly determines task allocation in automation systems:

Confidence above 90%: Execute automatically. Confidence between 70% and 90%: Route to a stronger large model for review. Confidence below 70%: Escalate to a human for decision-making.

Armin Ronacher, CTO of Earendil, explained: "Fundamentally, this approach shifts some responsibility for handling hallucination issues onto the user. The user must decide: a 50% probability might mean it's okay to ignore, but a 95% probability means I can use it."

Parallel Processing: Simultaneous Multi-Question Analysis

Jev offers another unique feature: multiple questions within a single request share one input but are evaluated independently and in parallel. A test by the public account "数字生命卡兹克" demonstrated this: feeding one news item allowed simultaneous questions - is it AI-related? Is it an ad? Is it fundraising? Which category fits? Is it worth pushing? Jev completes all these judgments at once without impacting speed.

Traditional large language models either generate responses sequentially or require bundling all questions into one prompt with output parsing. This parallel capability provides the most significant advantage in multi-question scenarios.

Real-World Applications by Developers

Since Jev's launch, the developer community has produced numerous practical implementations.

Feed filtering: Developer Marcel Pociot built a browser extension using Jev to filter unwanted content on X, with each post judgment averaging 380 milliseconds.

Ad analysis: One developer analyzed 724 live ads in 40 seconds, making 8,724 judgments total.

Browser agent: The open-source project jev-ultrafast by Browser Use employs Jev as a decision layer for browser operations. For flight searches, each round asks multiple questions simultaneously: click or type next? Which numbered element? All questions share the same webpage state while providing individual answers.

Context compression: Integrating Jev into Claude Code to identify critical context information enables near-instantaneous context compression.

Security classification: Pranit Sharma, a software engineer at Vercel, reported replacing OpenAI's ChatGPT Luna 5.6 with Jev for a command safety classifier, achieving 5 to 18 times faster processing with improved accuracy.

Email classification: Nikhil Mudholkar, CTO of Bryo AI, tested Jev and Gemini on business email classification. His results showed Gemini had slightly better accuracy but cost 10 to 20 times more.

Marketing analytics: One team integrated Jev into marketing analysis pipelines for scanning Meta's ad library and comparing ad format lifespans, compressing tasks that previously required human effort into 30x faster workflows, all under $3 in costs.

Gaming demonstration: TypeSafe's site showcases Jev playing Doom, making roughly 10 decisions per second based on structured game state, at approximately $7 per hour. In the Ably Pong demo, Jev made 47 operational decisions in 12 seconds, while Gemini, Claude, and GPT managed only two or three in the same timeframe.

Understanding the "Zero Hallucination" Claim

Jev's site prominently displays "Zero Hallucinations." While attention-grabbing, this claim is easily misunderstood. TypeSafe's technical documentation clarifies: Jev's output types and structures are predefined, so the model cannot produce type errors, deviate from specified formats, fabricate nonexistent fields, or expand beyond its defined options. For instance, if restricted to choosing among "pass, reject, human review," it cannot output a poem or append a Markdown explanation.

This addresses structural reliability in automation systems. It does not guarantee the model never makes incorrect choices. TypeSafe's own FAQ explicitly acknowledges that Jev "may still make mistakes."

Why Jev's Popularity Is Happening Now

Jev's emergence aligns perfectly with a practical problem in agent development. An agent capable of browsing web pages, using tools, and processing files must make numerous small decisions before completing a task. Which element to click next? Is this tool output valid? Which part of the context is now irrelevant? Is this task genuinely complete or does it only appear so?

These individual judgments seem trivial, but in high-frequency workflows, their costs accumulate rapidly. Jev's explosive popularity stems from three clear factors:

First, its positioning is refreshingly contrarian. While the industry chases longer reasoning chains, stronger chat capabilities, and expanded context windows, Jev abandons text generation entirely, focusing solely on selection, classification, scoring, and routing.

Second, it directly addresses the cost problem in agent operations. For systems making hundreds of thousands of micro-decisions daily, a 20–200x speed differential and 40–400x cost differential - even if only partially realized - is enough to draw developer interest.

Third, it makes "uncertainty" part of the interface. Previously, models always provided answers; the real headache for automated systems was that models often didn't know when they were wrong. Jev's solution: return not just the conclusion but also the measure of confidence in that conclusion.

The Name Behind the Technology: System One

TypeSafe categorizes Jev under "System One Models," a name drawn from Daniel Kahneman's "Thinking, Fast and Slow," distinguishing System 1 (fast, intuitive judgment) from System 2 (slow, deliberate reasoning). Jev clearly bets on the former.

TypeSafe doesn't position Jev as a replacement for all large models. Official use cases include: classification, routing, scoring, information extraction, workflow branching, content moderation, model output validation, and batch processing on large datasets. Complex reasoning, long-form generation, coding, and creative content remain the domain of traditional large language models.

In this division of labor, Jev handles high-volume low-latency decisions first; when confidence is insufficient, it escalates to slower, costlier models better suited for generation and reasoning, or to human intervention.

Funding and Accessibility

TypeSafe AI was co-founded by Diogo Almeida with Erik Gafni and Sasha Sheng, operating in stealth for approximately two years. The company officially launched on September 15, 2026, announcing a $40 million seed round led by DCVC.

TypeSafe indicates they are currently the only company pursuing this approach. Ronacher anticipates imitators will emerge as the practical value of this model gains validation. When asked whether the company considers itself a frontier lab, Almeida said: "The primary output of frontier labs tends to be either fear or hype. I hope our primary product is wisdom. We don't belong to that category of lab obsessed with creating infinite wealth, building religious narratives, or trying to manufacture gods in data centers."