AI Companies Develop Decision Models to Avoid Hallucinations

TypeSafe AI released a model called Jev on 15 September, billing it as a "System One model" that is cheaper and faster than large language models and, in the company's words, cannot hallucinate. The startup has raised $870 million at a $7.5 billion valuation. Within roughly six or seven weeks, according to independent developer Matt Mastracci, dozens of decision models appeared: Cloudflare, Amazon, OpenAI and Microsoft each shipped one, alongside start-ups, academic groups and hobbyists. Mastracci's own Djev was cited by Cloudflare as an inspiration for its models.
The System One label borrows from Daniel Kahneman's Thinking, Fast and Slow, which separates fast instinctive thought from slow deliberate reasoning. TypeSafe AI's argument is that LLMs supply the slow mode and are too slow for the fast one. The company claims end-to-end latency of 70 to 500 milliseconds for Jev, against at least several seconds for typical LLM responses and hours for heavy tasks such as writing software from scratch.
Independent testing broadly supports the speed claim. A pre-print co-authored by Ramayya Krishnan of Carnegie Mellon University and colleagues measured a median latency of 150 milliseconds for Jev across decision-accuracy benchmarks including RewardBench and JudgeBench. GPT-6 Astra, OpenAI's most capable model at the time, came in at 1.89 seconds median. Speed translates into price: TypeSafe AI charges $0.042 per million input tokens and nothing for output tokens, while GPT-6 Astra costs $10 per million input and $50 per million output. Even GPT-6 Luna, the cheap fast variant, runs $0.10 and $0.50 respectively. Mastracci points to workloads LLMs cannot practically handle today, such as moderating a firehose of social posts or a very busy chat room. Michelle Chen, a group product manager at Cloudflare, estimates that around 80 percent of typical LLM tasks could be offloaded to a System One model at far lower cost and latency.
TypeSafe AI has disclosed little about how Jev achieves its speed, but the open-weights imitators are more transparent. Cloudflare's Clef, free on Hugging Face and built on Alibaba's Qwen, keeps the transformer architecture and still takes tokens in and out. The difference is in the output stage. As Chen describes it, an LLM generates text one token at a time, re-reading everything produced so far at each step until it emits an end token. A decision model does a single forward pass over the input and produces one output. The parts of the network that turn tokens into free text are altered so the answer is constrained. Zhongyu Wei, a professor at Fudan University and co-author of a benchmarking study that introduces an open-weight alternative called SII-Decision, describes the format: the model receives a question plus options such as A, B and C, must pick one, and returns probabilities with its choice.
In practice this is a return to classifiers, the models behind spam filters and face detection. Chen calls it almost a rebranding. The upgrade is generalization. Old classifiers were trained per domain; a decision model inherits an LLM's broad competence, so, in Mastracci's description, you can pick it off the shelf and point it at your problem. Chen notes that building a classifier previously meant collecting data, standing up an ML team and fine-tuning. Cloudflare now uses decision models for flagging spam pull requests in its GitHub documentation repo, catching spam and phishing email, spotting phishing domains, detecting prompt-injection attacks on LLMs, and triaging emails about internet backbone outages to judge whether its services are affected.
The no-hallucination claim holds up narrowly. LLMs hallucinate because they sample repeatedly across a huge vocabulary and emit many tokens, leaving room for a plausible but wrong continuation. A decision model only scores the predefined options, so it cannot invent an answer outside them. Krishnan cautions that removing output hallucinations does not remove decision errors. In his tests, Jev judged answers well when the call was readable from the prompt, but degraded when the judgment needed math or a logic puzzle. Wei's group found decision models slightly ahead of LLMs on document understanding, roughly level on tool use, and behind in expertise-heavy fields such as medicine, finance and law. Wei points to reasoning as a key factor, while noting it does not explain every gap and that training method matters.
The two model types need not compete. Krishnan's team routed Jev's low-confidence answers to GPT-6 for a second look; the pairing slightly beat GPT-6 alone at 41 percent of the cost. Wei expects such hybrid architectures to become common, with decision models calling an LLM for explanation and LLMs calling decision models for quick judgments. My reading is that the economics favor exactly this split: cheap deterministic gating up front, expensive generation only when confidence is low or prose is required. The catch is that confidence calibration, not raw accuracy, becomes the critical engineering problem, and neither the pricing nor the marketing claims address it.
Why it matters: Teams building high-volume classification, moderation, triage or routing features now have a credible option that is orders of magnitude cheaper per call than a frontier LLM, and open-weights versions remove vendor lock-in. The trade-off is scope: these models choose among options you define and lag on expert-domain questions, so expect adoption as a first-pass filter in front of an LLM rather than a replacement, and plan evaluation around decision accuracy under uncertainty, not just absence of hallucinated text.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
On 15 September, a startup called TypeSafe AI released a new AI model called Jev . The company’s broad, disruptive claims about Jev that turned heads: The company claimed that Jev, which it calls a “System One model” is cheaper, faster, and can’t hallucinate. TypeSafe AI’s marketing is an obvious shot across the bow of other AI companies and their large language models (LLMs), which, despite their capabilities in math and coding, are infamous for their cost and tendency to “hallucinate” factually incorrect replies.