The clearest definition comes from the llama.cpp project, which added support for the format on 2 October: “A decision model reads the input once, and its answer is always one of your options, with a probability attached.”

That sentence separates it from the language models most people use. A language model generates text one token at a time, so using its judgement inside software means prompting for a structured answer and then parsing text that can come back malformed. A decision model takes an input once and scores the options it is handed, returning one of those options or a probability rather than prose: the caller declares the set in advance, and the answer is a probability that a condition is true, a choice among supplied values, or a score on ordered levels. There is no free text to get wrong in the hallucination sense, and no output type error, because the possible answers were fixed before the model ran. The risk that remains is choosing the wrong option from the set, sometimes with high confidence.

Mechanically, the pattern is the model’s hidden states feeding a small scoring head in place of a text-generation head, which makes the class closer to a classifier than to a new kind of intelligence.

Vendors pitch it for routing, classification, prioritisation, verification and workflow control. Microsoft describes decision models as “purpose-built to deliver structured outputs that software can immediately act on.” AWS, which shipped its own entry, describes the trade-off plainly: decision models are faster and stronger at a given size and always answer from the options, but they are “significantly worse at solving complex problems than reasoning models” and cannot generate text. AWS says they do not replace language models for coding, chat and summarisation.

The three entries shipped within four weeks

Jev came first. TechCrunch dates TypeSafe’s launch to 15 September, and The Register described it on 10 October as announced “three weeks ago.” OpenAI put its Decisions API into public beta on 6 October. Microsoft announced Microsoft-Decision-1 on 9 October, three days after OpenAI.

OpenAI’s Decisions API

OpenAI’s developer documentation lists gpt-6-luna as the only model on the endpoint, POST /v1/decisions, and three question types: predicate, which returns a probability between 0 and 1 that a condition is true; choice, which returns one supplied value with a probabilities array and a confidence field; and score, which returns a probability-weighted average of ordered levels. Images must be passed as inline base64, and hosted URLs and file IDs are not supported. Pricing is $0.10 per million input tokens, and OpenAI says it charges only for input tokens, with no output-token charges. OpenAI describes the endpoint as “about 10x faster than the Responses API” and says general availability is expected “in the coming weeks.” Both are OpenAI’s claims about its own product. The public beta began with a community announcement on 6 October.

Microsoft-Decision-1

Microsoft announced its “decision-scoring model” on 9 October, in a post by Achint Srivastava, a vice president of software engineering in the Office of the CTO. It is available in Microsoft Foundry and through OpenRouter. Microsoft’s own benchmark comparison, which it says covers 36 benchmarks and nearly 150,000 questions kept blind from training, puts Microsoft-Decision-1 at 83.5% average accuracy and 85 ms median latency, with a 95th-percentile latency of 125 ms. Microsoft calls it the “fastest measured,” claiming it is 2.5 times quicker than the runner-up, H2O-Lightning-4B v1.1, and 35 times quicker than GPT-6 Sol. The Register reported further figures from the same claims: 2.8 times faster than Jev in a latency test, more than 20 times cheaper than GPT-6 Sol on text classification, and input at $0.042 per million tokens with output free. Microsoft’s own chart gives Microsoft-Decision-1 a calibration score of 92.2, third behind Jev at 93.7 and Quyet-1.0-Large at 93.1, and its note reads “Third on calibration, 1.5 behind Jev.” Every one of those numbers comes from Microsoft’s own testing.

The provenance of the first version is worth stating. The Register reports that the first Microsoft-Decision-1 is based on Alibaba’s Qwen3.5-9B, an existing open-weight model, and that Microsoft says a later version will be rebased on its own models and OpenAI’s. Microsoft’s announcement post does not mention that provenance.

TypeSafe’s Jev

Jev is a “System One model,” in TypeSafe’s language, and the company’s documentation describes three primitives: choice and score return probabilities and a confidence value, and noul returns a single probability between 0 and 1. TypeSafe says Jev is built on a new architecture with a parallel sampler and a training method it calls Reinforcement Learning for Calibrated Decisions. It lists input at $0.042 per million tokens with output free, response times of 70 to 500 ms, and claims it is “two orders of magnitude faster and more efficient” than the alternative and that it “can’t hallucinate,” in its launch post. The company announced an $870 million funding round at a $7.5 billion valuation on 9 October, led by Andreessen Horowitz with Sequoia Capital and DCVC participating, in its Series AI announcement.

Other vendors shipped comparable models in the same stretch, among them Cloudflare’s Clef, AWS’s Strands Decider 2B and Liquid AI’s d1, and The Register reported “more than 100 such models” in the field on 10 October.

Every performance figure the vendors published comes from the vendor selling the model

There is no independent head-to-head benchmark comparing OpenAI’s Decisions API, Microsoft-Decision-1 and Jev on a shared task set. Microsoft’s 36-benchmark result is run and chosen by Microsoft. OpenAI’s “10x” is OpenAI’s. The community leaderboard several vendors quote, known as JevBench and published as the Decision Index, could not be read directly for this article because its page loads no data without JavaScript, so its rankings appear only as vendor-quoted positions.

Confidence and calibration scores are not comparable across the three products. OpenAI returns a confidence value for choice questions and a probability for predicate questions. Microsoft publishes a calibration score, where 100 means confidence matches how often the model is right. Jev publishes its own confidence. Each is built over different tasks and defined differently, so a sentence claiming one model is more confident or more accurate than another cannot be supported by these numbers alone.

TypeSafe’s adoption and savings lines are marketing. “A third of the Fortune 500 are getting their Jev on” and “We’ve saved customers millions of dollars in production already” appear on the company’s funding announcement with no customer list, deployment count or methodology. TechCrunch reported the Fortune 500 line as something “the startup claims.” Neither figure should be read as verified adoption or savings.

User reports hint that the answer depends on how you ask

Separately from vendor benchmarks, a thread on OpenAI’s community forum from 6 and 7 October describes user-reported behaviour to treat as anecdote, not measurement. One user reported that a biased coin that should land heads 70% of the time returned about 70% heads under predicate framing over 1,000 trials, but about 98% heads when the same question was framed as a choice. Changing the order of the options also changed the probabilities, and the same user’s conclusion was to avoid choice questions. Another user posted an informal comparison of Luna and Jev that favoured Jev on judged tasks and cost, which they labelled “a strong early signal rather than a general benchmark.” These are self-chosen tests with no methodology control.

One independent study found confident answers can be redirected

An independent academic study, JevOut, reports that among decisions each system initially answered correctly, short additions to the context that preserve the question and the correct answer redirected 61.4% to 73.2% of decisions to a wrong option chosen in advance, including decisions the model had answered with high confidence. Those rates are the share of each system’s initially correct decisions for which an automated, probability-guided search, within 64 accepted attempts per decision, found at least one redirection, and a single added sentence sufficed for 211 of 508 Jev decisions. The authors state the rates are “not a universal failure probability for all model inputs,” and they distinguish the work from prompt injection. The study is listed as arXiv 2609.30243 and accepted as a poster at Agenthon 2026 at NeurIPS 2026. It studies decision reliability rather than downstream harm, and it does not execute tools. Among the evidence gathered for this article, it is the strongest independent counterevidence found, and it qualifies any claim that these models cannot be wrong.

The class is real, the label is not yet settled

Several independent vendors shipped a recognisable shape in the same weeks: an input in, a probability over caller-supplied options out. Two platform vendors did so within three days. That convergence is real.

Whether “decision model” settles as a category name is a different question. The naming is contested and overlaps with older ideas, including ordinary classifiers and the decision-model notation used in business rules. Every vendor coins its own vocabulary: OpenAI sells the Decisions API with predicate, choice and score questions; Microsoft says “decision model” and “decision-scoring”; TypeSafe and llama.cpp both use the System One format, with choice, score and noul, and llama.cpp deliberately follows TypeSafe’s format while OpenAI uses its own field names. There is no single shared standard behind the shared shape. Readers will meet the term whether or not it lasts.