New Jev model answers in 150 ms versus 1.89 s for GPT-6 Astra
If you just need to decide quickly whether an email is spam or a comment breaks the rules, you don't need a large LLM to write out text. Startup TypeSafe AI has released Jev, a model that picks one option from a given set (A, B or C) and responds in 150 ms on average, while GPT-6 Astra takes 1.89 seconds. It cannot invent an answer outside the list, but it still makes mistakes on math and logic problems. That makes it a fit for simple verdicts, while harder tasks should stay with a large LLM.
AI-processed from IEEE Spectrum AI; edited by Hamidun News
If you run comment moderation, phishing checks or ticket triage through a large language model, you are probably waiting longer and paying more than you need to. These tasks don't call for coherent prose. A short verdict is enough: "yes," "no" or "option B." That is the idea behind a new class of "decision models." On September 15, startup TypeSafe AI released the best-known of them, Jev. According to an independent evaluation, its median latency was 150 milliseconds, while OpenAI's GPT-6 Astra responded in 1.89 seconds. Let's look at where that gap comes from and where it stops.
Why decision models are fast
TypeSafe AI calls Jev a "System 1 model," a reference to Daniel Kahneman's book Thinking, Fast and Slow, which contrasts fast, intuitive thinking with slow, deliberate thinking. In the company's view, ordinary LLMs deliver the second kind but are too slow for the first. Most researchers simply call such models decision models.
The mechanics are simple. An LLM generates an answer one token at a time: each new token depends on all the previous ones, until the answer ends. As Michelle Chen of Cloudflare explains, a decision model doesn't work that way. It reads the entire input in a single pass and immediately returns one result. Under the hood it is still a transformer (Cloudflare's Clef, for example, is built on Alibaba's Qwen), but the output is tightly constrained. You supply a question and options A, B and C, and the model returns a choice along with a probability for each option.
The company claims an end-to-end latency of 70–500 ms. A Carnegie Mellon group led by Ramayya Krishnan measured a median of 150 ms on the RewardBench and JudgeBench tests. The pricing is different, too: Jev charges $0.042 per million input tokens, and output tokens are free. GPT-6 Astra costs $10 for input and $50 for output, and the lighter GPT-6 Luna costs $0.10 and $0.50. Investors bought the pitch: TypeSafe AI raised $870 million at a $7.5 billion valuation, and Cloudflare, Amazon, OpenAI and Microsoft quickly released their own decision models.
Old classifiers with a new skill
Chen admits this is close to a rebranding: classifiers such as spam filters and face recognition have existed for a long time. But previously, each task meant collecting data, assembling a machine learning team and fine-tuning a model. A decision model inherits the general knowledge of an LLM, so, in the words of independent developer Matt Mastracci, you can "take it off the shelf, point it at your task, and it works."
Cloudflare already uses this approach. It filters out spam pull requests in documentation repositories, catches spam and phishing emails and domains, detects prompt injection attacks on LLMs, and reads emails about backbone link outages to work out whether they will affect the company's services. Mastracci sees potential in moderating a social media feed or a very busy chat.
No hallucinations, but there are errors
TypeSafe AI's claim that Jev cannot hallucinate is technically correct. LLMs make things up because at every step they choose from a huge vocabulary and can veer off course. A decision model calculates probabilities only among predefined options and picks the most likely one, so an answer outside the list is impossible.
But Krishnan warns that having no hallucinations in the output does not mean having no errors in the decision. Jev did well at judging whether an answer to a question was correct when that was evident from the text itself, but it did worse when the task required math or a logic puzzle. A study by Zhongyu Wei of Fudan University, which presents the open model SII-Decision, showed a similar picture: the best decision models are slightly more accurate than LLMs at document understanding and close to them at picking the right software tool, but they fall behind where domain expertise is required.
What to do in practice
Chen believes that up to 80% of typical LLM tasks can be shifted to a decision model. That is a vendor's estimate, and it should be treated as a hypothesis. It is easy to test: take your request log, pull out the "choose an option" tasks and run them through the open Clef or SII-Decision. If quality on your sample is comparable, you will gain in speed and cost. If not, keep those cases on a large model.
- Nov 18, 2026AWS adds Dynamo to HyperPod: Qwen3-235B delivers 207 tokens/s per GPU instead of 101
- Nov 18, 2026Huang steps down as NVIDIA CEO, company loses more than $1 trillion in market cap
- Nov 18, 2026The Best Coding Agent Controls a Robot in 77% of Tasks, but With Sensors Alone Only 30%
- Nov 18, 2026Moonshot AI Shifts Kimi From Code to Science: Gene Analysis in 90 Minutes Instead of 4 Hours
Key Questions Answered
Why decision models are fast?
TypeSafe AI calls Jev a "System 1 model," a reference to Daniel Kahneman's book Thinking, Fast and Slow, which contrasts fast, intuitive thinking with slow, deliberate thinking. In the company's view, ordinary LLMs deliver the second kind but are too slow for the first. Most researchers simply call such models decision models. The mechanics are simple. An LLM generates an answer one token at a time: each new token depends on all the previous ones, until the answer ends.
What are the key details on Old classifiers with a new skill?
Chen admits this is close to a rebranding: classifiers such as spam filters and face recognition have existed for a long time. But previously, each task meant collecting data, assembling a machine learning team and fine-tuning a model. A decision model inherits the general knowledge of an LLM, so, in the words of independent developer Matt Mastracci, you can "take it off the shelf, point it at your task, and it works." Cloudflare already uses this approach.
What are the key details on No hallucinations, but there are errors?
TypeSafe AI's claim that Jev cannot hallucinate is technically correct. LLMs make things up because at every step they choose from a huge vocabulary and can veer off course. A decision model calculates probabilities only among predefined options and picks the most likely one, so an answer outside the list is impossible. But Krishnan warns that having no hallucinations in the output does not mean having no errors in the decision.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.