Decision Models Go Mainstream: Faster, Cheaper, and Hallucination-Free?

TypeSafe AI's Jev enters the market with sub-second latency and ultra-low pricing, but "hallucination-free" does not mean "error-free."
TypeSafe AI launched the decision model Jev, marketed on sub-second latency, ultra-low pricing, and being "hallucination-free," quickly prompting Cloudflare, Amazon, OpenAI, and Microsoft to follow suit, and raising $870 million. Decision models retain the LLM's transformer architecture but perform only a single forward pass and constrain output to a predefined structured format—essentially classifiers with generalization ability. Independent research confirms their speed advantage while noting that "eliminating output hallucination" is not the same as "eliminating decision errors": in math, logic, and specialized domains such as medicine, finance, and law, decision models still lag behind LLMs. Researchers therefore favor hybrid architectures combining decision models and LLMs—using the former for high-frequency fixed judgments and the latter for reasoning and specialized knowledge—to balance accuracy and cost.

Large language models (LLMs) shine at math and coding, yet they are widely criticized for their high cost and their tendency to "hallucinate" factual errors. A startup called TypeSafe AI claims that its Jev model is cheaper, faster, and "does not hallucinate." That claim quickly stirred up the entire industry—Cloudflare, Amazon, OpenAI, and Microsoft have successively rolled out "decision models" imitating Jev's capabilities, while independent developers and academic researchers have contributed dozens of variants.
Why Jev Is Shaking Up the Market
TypeSafe AI calls Jev a "System One model," borrowing from the cognitive theory psychologist Daniel Kahneman laid out in Thinking, Fast and Slow: System One handles fast, instinctive thinking, while System Two handles slow, deliberate thinking. The company's argument is that modern LLMs provide System Two but are too slow to take on the role of System One.
Independent testing corroborates the speed claims. A preprint paper co-authored by Ramayya Krishnan, a professor of information systems at Carnegie Mellon University, found that Jev's median latency on benchmarks such as RewardBench and JudgeBench—which measure decision accuracy—was far lower than that of its LLM competitors at the time. OpenAI's then-strongest model, GPT-6 Astra, had a median latency in the seconds, whereas Jev sat in the sub-second range. TypeSafe AI's own marketing puts end-to-end response latency within tens to hundreds of milliseconds, in contrast to LLMs' multi-second responses and hours-long runtimes on complex tasks.
Faster replies usually mean lower cost. TypeSafe AI charges an extremely low fee for input tokens and offers output tokens for free; by comparison, GPT-6 Astra's input and output pricing is several orders of magnitude higher, and even GPT-6 Luna—positioned for fast, cheap inference—is noticeably more expensive.
On the capital side, the company recently closed an $870 million funding round at a valuation of $7.5 billion. Independent developer Matt Mastracci remarked that Jev "really brought decision models to the forefront," with dozens of models emerging in just six or seven weeks.
Daniel Kahneman's dual-system theory comes from his 2011 book Thinking, Fast and Slow. System One refers to automatic, intuitive, low-energy cognitive processes—such as recognizing an angry face or answering 2+2; System Two requires attentional effort and handles complex calculation, logical reasoning, and self-control. Kahneman used this framework to explain many biases in human judgment: System One quickly offers an intuitive answer, and System Two, which ought to review it, finds the cost of correction too high—so people often accept the wrong intuition. TypeSafe AI borrows this metaphor to argue that LLMs play the role of "slow thinking," while a great many real business scenarios (such as email classification, content moderation, and routing decisions) only need "fast thinking."
RewardBench and JudgeBench are benchmarks for evaluating a model's ability to act as a "judge." RewardBench measures a reward model's accuracy in judging conversational preferences, while JudgeBench examines a model's ability to correctly adjudicate the quality of responses. What they share is that the task format is fixed and the output is a limited set of options—precisely the sweet spot for decision models. This also explains why Jev's latency advantage on these benchmarks translates directly into a usability advantage.
How Decision Models Work
Decision models did not appear out of nowhere. Cloudflare open-sourced Clef on Hugging Face, built on Alibaba's Qwen LLM. According to Cloudflare product manager Michelle Chen, an LLM generating output looks at all previous tokens one token at a time and predicts the next, whereas a decision model performs a single forward pass—reading all the input and directly producing one output—which is exactly why it is so much faster.
These models still retain the neurons and transformer architecture of LLMs, and still take tokens in and out, but the part that turns tokens into text is constrained. Research involving Zhongyu Wei, a professor at Fudan University, points out that Jev can only answer in a predefined structured format: given a question and options such as A, B, and C, it must choose among them and return a probability at the same time.
In a sense, decision models are a return to classifiers. Classifiers have long been used for spam filtering and face detection. Chen put it bluntly: "These models aren't new—you could almost call it a repackaging."
The real upgrade lies in generalization. Mastracci explains that traditional classifiers need to be trained for a specific domain, whereas decision models retain the general intelligence of an LLM and can be "used out of the box—point them at your problem and they work." Cloudflare uses this capability in a variety of scenarios: flagging spam pull requests in GitHub documentation repositories, blocking spam and phishing emails, identifying phishing domains, detecting prompt injection attacks against LLMs, and reading emails about internet backbone outages to judge whether they will affect its own services.

Forward pass refers to a neural network completing one full computation from the input layer to the output layer, without involving backpropagation or iterative generation. An LLM's autoregressive decoding requires repeating a forward pass for each output token—generating a response of length N means running it N times; a decision model runs it only once, reading all the input and then outputting a probability distribution over a preset candidate space. This difference is the fundamental source of the order-of-magnitude speedup, rather than mere model compression or hardware optimization.
Prompt injection is an attack against LLM applications: an attacker embeds instructions in input content (such as web pages, emails, or documents) to induce the model to ignore its original system prompt and carry out the attacker's intent. Cloudflare uses decision models to detect such attacks because, once "is this text trying to override the agent's instructions" is turned into a binary classification or probability scoring task, a single forward pass suffices—and the output is constrained to a limited set of labels, so it won't be "led astray" by malicious text the way a generative model can be.
The Truth and Limits of "Hallucination-Free"
The name TypeSafe itself is an analogy: a type-safe language rejects incompatible data operations before the program runs, rather than producing bad results afterward; decision models likewise prevent the model from generating unexpected replies.
Strictly speaking, "Jev never hallucinates" holds true. LLMs hallucinate because they must repeatedly compute token probabilities over a vast vocabulary and output a large number of tokens; a low-probability token can derail the response, and the model may output tokens that are plausible under its training distribution but inconsistent with the context. A decision model computes probabilities only among predetermined possibilities and selects the most likely one, so hallucination is eliminated.
But Krishnan cautions: "Eliminating hallucination in the output does not equal eliminating errors in the decision." He and his co-authors had Jev judge whether answers to questions were correct, and found that it performed well when the judgment could be made simply by reading the prompt, but that its performance dropped noticeably once math or logic puzzles were involved.
Wei's team reached a similar conclusion: the best decision models can slightly outperform LLMs on document-understanding questions and come close to LLMs on tool-use questions (for example, choosing the right software tool to complete a user's task), but they lag behind in domains requiring specialized knowledge such as medicine, finance, and law. Wei believes reasoning ability is one key factor—reasoning lets an LLM sort out its thinking before giving a final answer—but reasoning cannot explain all the domain gaps, and the method used to train decision models may also be a variable; that is why they adopted a training approach tailored to decision tasks in SII-Decision.
Type safety is a programming language concept: the compiler statically checks before the program runs whether the types of variables and operations match, rejecting incompatible combinations at compile time rather than waiting for a runtime crash or erroneous result. TypeSafe AI's company name and product positioning both stem from this analogy—treating "the model's output must be one of a predefined set of structures" as a kind of type constraint that mechanically rules out the possibility of generating illegal output. Note that type safety guarantees that "the program won't fail due to type errors," not that "the program's logic is correct"; likewise, a decision model guarantees that the output format is valid and falls within the candidate set, but does not guarantee that the chosen option is semantically correct. This is precisely the technical root of what Krishnan calls "eliminating hallucination is not the same as eliminating decision errors."
Hybrid Architectures May Become the Norm
Decision models and LLMs are not necessarily competitors. Krishnan and his colleagues also tested passing Jev's low-confidence answers to GPT-6 for further processing, and found that this combination was slightly better in accuracy than using GPT-6 alone, at significantly lower cost.
Wei believes this kind of "hybrid architecture" will become common: "A decision model can call an LLM to provide a more detailed explanation, and an LLM can call a decision model to make a quick judgment. I think this is very flexible."
For developers, this means they don't have to make an either-or choice between speed and accuracy. Handing high-frequency, fixed-format judgments to decision models and leaving tasks that require reasoning and specialized knowledge to LLMs may be the more pragmatic engineering path today.
One-line Summary
TypeSafe AI's Jev enters the market with sub-second latency and ultra-low pricing, but "hallucination-free" does not mean "error-free."
Paragraph Summary
TypeSafe AI launched the decision model Jev, marketed on sub-second latency, ultra-low pricing, and being "hallucination-free." It quickly prompted Cloudflare, Amazon, OpenAI, and Microsoft to follow suit, and the company closed an $870 million funding round. Decision models retain the LLM's transformer architecture but perform only a single forward pass and constrain output to a predefined structured format—essentially classifiers with generalization ability. Independent research confirms their speed advantage while noting that "eliminating output hallucination" is not the same as "eliminating decision errors": in math, logic, and specialized domains such as medicine, finance, and law, decision models still lag behind LLMs. Researchers therefore favor hybrid architectures combining decision models and LLMs—using the former for high-frequency fixed judgments and the latter for reasoning and specialized knowledge—to balance accuracy and cost.
Related articles

Why Git Is Essential in the AI Era: From Version Control to Agent Collaboration
What's the difference between Git and GitHub? How do you manage changes with Issues, branches, and Pull Requests after AI Agents write code? Based on NetworkChuck's Git tutorial, this article walks through the complete flow from git init to conflict resolution and analyzes why version control has become an essential skill in the AI era.

OpenAI 6.2 Astra Hands-On Review Gone Wrong: Impressive Benchmarks, But Developers Say It's Unusable
OpenAI 6.2 Astra leads benchmarks at lower cost, but early testers report frequent stalls in long tasks, unparseable output, and uncontrolled resource management. We examine the gap between real-world experience and official data, and what it means for developer adoption.

GTA VI Coffee System Details Revealed: Hot Coffee Can Be Splashed on People?
Vergecast podcast reveals GTA VI's coffee system: players can brew in apartments, buy from vending machines, order in restaurants, and even splash hot coffee on others. This detailed recreation of daily interactions continues Red Dead Redemption 2's immersive design and sparks discussion on game detail versus gameplay balance.