TypeSafe AI's Jev decision model draws imitators from Cloudflare, Amazon, OpenAI

TypeSafe AI's Jev decision model draws imitators from Cloudflare, Amazon, OpenAI

On 15 September, a startup called TypeSafe AI released Jev, a model it calls a 'System One model'. The company claims Jev is cheaper, faster and cannot hallucinate. The name comes from Daniel Kahneman's Thinking, Fast and Slow, where System One is fast, instinctive thought and System Two is slower, deliberate thought. TypeSafe AI argues that modern LLMs offer the second but are too slow for the first. Most AI researchers would call Jev a decision model or decider model. The startup recently raised $870 million at a valuation of $7.5 billion.

The article says the most impressive sign of Jev's success is the imitators. Cloudflare, Amazon, OpenAI and Microsoft quickly released decision models that imitate Jev, and start-ups, academic researchers and independent enthusiasts have released dozens more. Matt Mastracci, an independent developer who has trained and released several decision models including Djev (which Cloudflare cited as an inspiration), says Jev 'really brought decision models to the forefront' and that dozens of models came out in six or seven weeks.

On speed, TypeSafe AI claims end-to-end response latency between 70 and 500 milliseconds, against several seconds at minimum for typical LLMs. The article says independent testing backs this up. A pre-print co-authored by Ramayya Krishnan, a professor of information systems at Carnegie Mellon University, and several colleagues found a median latency of 150 milliseconds for Jev across several decision-accuracy benchmarks, including RewardBench and JudgeBench. OpenAI's most capable model at the time, GPT-6 Astra, had a median of 1.89 seconds.

Price follows speed. TypeSafe AI charges US $0.042 per million input tokens, and output tokens are free. ChatGPT charges $10 per million input tokens and $50 per million output tokens for GPT-6 Astra. Even GPT-6 Luna, a variant built for fast, inexpensive inference, costs $0.10 and $0.50 per million input and output tokens. Mastracci says the lower latency and cost open tasks LLMs cannot handle now, such as moderating a fire hose of social media posts or a very busy chat room. Michelle Chen, a group product manager at Cloudflare, says you can 'probably' offload 80 percent of typical LLM tasks to a System One model; that is her estimate, not a measurement.

TypeSafe AI has said little about how Jev gets its speed, but open-weights imitators are more transparent. Cloudflare's Clef, built on Alibaba's Qwen LLM and free to download on Hugging Face, does not generate token after token as an LLM does. Chen says it makes a single forward pass that reads all the inputs once and then produces one output. Such models keep an LLM's neurons and transformer architecture and still take and emit tokens, but the part that turns tokens into text is changed to constrain the output. Zhongyu Wei, a professor at Fudan University and co-author of a study that benchmarks decision models and introduces an open-weight alternative, SII-Decision, says Jev can only answer with a predefined, structured format: it is given a question and options such as A, B and C, and must pick one, returning probabilities alongside its choice.

In one sense decision models are a return to classifiers, the kind of model behind spam filters and face detection; Chen calls it 'almost like a rebranding'. The upgrade is generalization. Classifiers used to need domain-specific training, with lots of data and a machine learning team. Decision models keep an LLM's general intelligence, so, in Mastracci's words, you can pick one up off the shelf, point it at your problem, and it works. Cloudflare uses them to flag spam pull requests in its GitHub documentation repository, catch spam and phishing emails, spot phishing domains, detect prompt-injection attacks on LLMs, and read emails about internet backbone outages to judge whether they will affect its services.

The no-hallucination claim is, the article says, strictly speaking true. TypeSafe AI's name nods to type-safe programming, and the company says Jev will never hallucinate. LLMs hallucinate because they calculate probabilities across a huge vocabulary over many tokens, which lets improbable or contextually wrong tokens lead a reply astray. A decision model only calculates probabilities among pre-determined options and picks the most probable. But Krishnan cautions that eliminating hallucinations in the output does not eliminate errors in decision-making. He and his co-authors asked Jev to judge whether answers to questions were correct. It did well when the judgment could be reached by reading the prompt, and worse when the judgment needed math or a logic puzzle.

Wei's group reports similar results. The best decision models can slightly beat LLM accuracy on document-understanding questions and come close on tool-use questions, such as choosing the right software tool to call. They fall behind in fields needing topic expertise, including medicine, finance and law. Wei says reasoning is a key factor, though it does not explain every gap, since some LLMs did well without reasoning; training methods can matter too, and SII-Decision used decision-specific training.

The two kinds of model need not compete. Krishnan and colleagues passed Jev's low-confidence answers to GPT-6. The combination slightly beat GPT-6 alone on accuracy at 41 percent of the cost. Wei thinks this hybrid architecture will become common, with a decision model calling an LLM for a detailed explanation and an LLM calling a decision model for a quick judgment.

Key facts

  • TypeSafe AI released Jev on 15 September and claims it is cheaper, faster and cannot hallucinate; the startup recently raised $870 million at a $7.5 billion valuation.
  • TypeSafe AI claims 70 to 500 ms latency. A pre-print co-authored by Ramayya Krishnan measured a median of 150 ms for Jev against 1.89 s for GPT-6 Astra.
  • Jev costs US $0.042 per million input tokens with free output; GPT-6 Astra costs $10 input and $50 output per million tokens.
  • Cloudflare, Amazon, OpenAI and Microsoft quickly released imitating decision models, with dozens more from start-ups, academics and enthusiasts.
  • Jev passing low-confidence answers to GPT-6 slightly beat GPT-6 alone on accuracy at 41 percent of the cost, but Jev's performance dropped on math and logic puzzles.

Why it matters

Decision models are a new category of fast, cheap model that chooses among fixed options instead of writing free text. TypeSafe AI's Jev pushed the idea into the open, and big vendors followed within weeks. The speed and price gap is large: a 150 ms median against 1.89 s for GPT-6 Astra in the pre-print, and $0.042 per million input tokens with free output against $10 and $50. The article also notes that the idea is partly old, since decision models return to classifiers, with generalization as the key upgrade.

Who it affects

Teams that run high-volume judgment tasks are the clear audience: content moderation of social posts or busy chat rooms, spam and phishing filtering, and detection of prompt-injection attacks. Cloudflare already uses decision models for several of these. Developers who built classifiers by collecting data and training them for a single task can now use one model across many domains. Providers of large LLMs face a cheaper, faster competitor for part of their workload.

How to use it

Decision models suit quick judgments on emails, messages or documents, where you give a question and a fixed set of options and get a choice with probabilities back. Cloudflare's Clef is an open-weights decision model, free to download on Hugging Face, and Wei's group offers the open-weight SII-Decision. Jev is priced at US $0.042 per million input tokens, with output tokens free. The hybrid pattern from the testing is to let the decision model answer and pass its low-confidence cases to an LLM such as GPT-6.

How solid is it

Mixed but reasonably good on speed. The 70 to 500 ms range is TypeSafe AI's own claim, while the 150 ms median comes from a pre-print co-authored by Krishnan and colleagues, and the article says independent testing backs the claim up. The no-hallucination claim is true in the narrow sense that the model only picks among pre-determined options. Chen's 80 percent offload figure is a hedged estimate, not a measurement. TypeSafe AI has offered few details about how Jev achieves its speed; the single-forward-pass explanation was given for Cloudflare's Clef. The article gives no accuracy figure for Jev or the hybrid beyond 'slightly beat' GPT-6 alone.

Risks and caveats

Krishnan's warning is the central caveat: removing hallucinations does not remove errors in decision-making. Jev did well when the answer could be read off the prompt and worse on math or logic puzzles. Wei's research found decision models behind LLMs in domains that need topic expertise, including medicine, finance and law. Jev can only answer in a predefined, structured format. The 41 percent cost figure comes from the tested benchmarks, and the article does not say it holds beyond them. No release dates, names or pricing are given for the Amazon, OpenAI and Microsoft models.

“eliminating hallucinations in the output does not eliminate errors in decision-making.”

— Ramayya Krishnan, professor of information systems at Carnegie Mellon University