🔍 Read the full analysis: Could A Model That Doesn’t Write Sentences Still Be The Most Useful AI? on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
TypeSafe AI has launched Jev, a model that forgoes sentence generation, instead producing typed, structured decisions. This approach aims to improve automation speed, reduce costs, and address issues like hallucinations in enterprise AI.
On September 15, 2026, TypeSafe AI introduced Jev, a new class of AI model that does not generate natural language sentences but instead provides structured, typed decisions with associated probabilities. This development challenges the prevailing assumption that large language models (LLMs) must produce text to be useful, especially within enterprise automation, and could significantly impact how AI is integrated into business workflows.
Jev is built for decision-making inside software systems, delivering structured answers such as ‘team: billing, confidence: 0.94’ rather than prose. It is designed to handle questions of choices, scores, and yes/no probabilities, acting more like a function call than a conversational agent. The model is optimized for speed, answering in 70 to 500 milliseconds at a cost of approximately $0.042 per million tokens, claims TypeSafe, which is hundreds of times faster and cheaper than traditional LLMs.
Developed by Diogo Almeida, a co-inventor of RLHF (Reinforcement Learning with Human Feedback) at OpenAI, Jev employs a training method called Reinforcement Learning for Calibrated Decisions (RLCD). TypeSafe argues that RLHF creates issues such as mode dropping and overconfidence, necessitating an alternative approach. Jev’s design aims to eliminate hallucinations related to output formatting errors, though it does not address errors stemming from incorrect decisions among allowed options.
In initial benchmarks, Jev tied with leading models at approximately 67.8% agreement when compared to GPT-6 Astra and Claude Fable 5.1, but independent tests show it scored around 62.6% on a phishing email detection task, with overconfidence issues on some question types. The company emphasizes that accuracy depends on how the judgment is broken into smaller questions and combined, not on the model alone.
Jev vs. LLMs: who should make the call?
Jev, from TypeSafe AI, is a “System One” model. It doesn’t write text. It returns a typed decision with a confidence score that your software can act on directly.
Same support ticket, two kinds of answer
“This ticket appears most likely related to billing, although it could also concern account settings or a recent plan change. I would suggest reviewing the invoice history before…”
A person reads it, or code has to parse the prose.
team: "billing"Software reads it and acts. Nothing to parse.
How they differ
| LLM | Jev | |
|---|---|---|
| Output | Text written for people | A choice, a score or a yes/no probability |
| Speed | Seconds per call | 70–500 ms* |
| Price | Input and (pricier) output tokens | $0.042 per million input tokens, output free* |
| Knows when it’s unsure | Often sounds confident when wrong | Confidence score on every answer |
| Explains its answer | Yes | No, which matters for audits |
| Best at | Reasoning, writing, open questions | Routing, tagging, scoring, duplicate checks |
* Vendor-reported. TypeSafe also claims up to 194× faster and 445× cheaper on its own selected workflows.
Accuracy is something you build
Jev is far cheaper and faster, but not more accurate than frontier models. How you phrase the question matters a lot.
TypeSafe’s benchmark scores agreement with two frontier models rather than verified ground truth. The five-question result used weights fitted on 1,000 labelled examples.
The real idea: a confidence dial you control
“duplicate listing”, confidence 0.62
Raise the threshold for fewer mistakes and more manual review. Lower it for more automation and more risk.
Only use Jev when all four hold
Good fits
- Routing tens of thousands of support tickets a day
- Flagging duplicate listings in a product catalogue
- Replacing a keyword filter that mis-tags half its matches
Poor fits
- Drafting customer emails or release notes
- Reviewing a few high-stakes contracts a month
- Anything that needs a written explanation
Implications for Enterprise AI and Automation
The introduction of Jev signals a potential shift in enterprise AI development, emphasizing structured decision outputs over natural language, which could lead to faster, more reliable, and cost-effective automation. By reducing reliance on text generation, companies may address common issues like hallucinations and formatting errors that cause pipeline failures. This approach aligns with the broader trend of integrating AI more deeply into operational systems, where predictable, schema-compliant outputs are critical.
Moreover, Jev’s speed and low cost could expand automation possibilities into areas previously limited by the expense and latency of traditional LLMs. As many enterprise decisions are simple judgments, this model could replace human-in-the-loop reasoning for routine tasks, freeing up resources and increasing efficiency. However, questions about its accuracy and reliability remain, especially in complex or ambiguous scenarios.
enterprise decision automation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on AI Decision-Making Models
Over the past three years, the AI industry has focused heavily on improving large language models like GPT and Claude, aiming for better reasoning, longer context, and more sophisticated code generation. These models produce natural language, which has proven useful for chatbots, content creation, and general-purpose AI applications. However, issues such as hallucinations, overconfidence, and high costs have limited their deployment in critical enterprise workflows.
In response, some researchers and companies have questioned whether text generation is always necessary or optimal. Diogo Almeida’s work at OpenAI on RLHF, which trains models to produce more human-aligned outputs, has been influential. Now, with Jev, TypeSafe is exploring a different paradigm: structured, typed decisions that can be directly integrated into software systems, bypassing the need for language-based reasoning.
“Jev is designed to produce structured decisions instead of text, making automation faster, cheaper, and more reliable in enterprise settings.”
— Diogo Almeida, CEO of TypeSafe AI
Limitations and Accuracy Challenges of Jev
While Jev shows promise, its accuracy benchmarks are based on agreement with frontier models rather than definitive correctness. Initial tests indicate a performance of around 62.6% on phishing detection tasks, which suggests room for improvement. Additionally, overconfidence in some answers and underconfidence in others raise questions about its reliability in high-stakes scenarios. It remains unclear how Jev performs across diverse, real-world enterprise tasks and how it handles ambiguous or complex judgments.
Expected Developments and Validation Efforts
TypeSafe plans to conduct broader independent evaluations of Jev in various enterprise applications to validate its accuracy and reliability. The company also intends to refine its training techniques and expand its benchmarking to better understand where Jev excels and where it needs improvement. As adoption grows, feedback from early users will inform further development, potentially leading to more robust, decision-focused AI models that complement or replace traditional LLMs in automation workflows.
Key Questions
How does Jev differ from traditional language models?
Jev produces structured, typed decisions with probabilities instead of natural language text, making it more suitable for automation and reducing errors related to formatting or hallucinations.
Can Jev replace human judgment entirely?
While Jev aims to automate routine decisions efficiently, its accuracy and reliability in complex scenarios are still being tested. It is likely to complement, rather than fully replace, human judgment in the near term.
What are the main limitations of Jev currently?
Its performance depends heavily on how questions are structured, and current benchmarks show room for improvement in accuracy. Overconfidence and handling ambiguous cases remain challenges.
Will Jev be suitable for all enterprise tasks?
Jev is designed for specific decision-making tasks that can be broken into small, well-defined questions. It may not be suitable for complex reasoning or creative tasks requiring nuanced understanding.
What is the significance of this development for AI research?
Jev represents a shift towards more deterministic, schema-based AI models, challenging the dominance of text-generation approaches and opening new avenues for reliable automation.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
