After Jev: the new decision models from OpenAI, Microsoft, Amazon and co. at a glance

A new category emerges within weeks

A decision model receives a situation and a clearly posed question, for example: which team is responsible for this request, billing, sales or retail? Instead of formulating an answer, it outputs a probability for each option. There are three question types: a choice among fixed options, a rating on a scale and a yes-or-no question with a probability. TypeSafe calls such models System One models, after the fast, intuitive thinking described in psychology.

TypeSafe introduced Jev on 15 September 2026. What happened next, Forbes describes as the birth of a category of its own: within a few weeks, large providers and independent developers released dozens of competing models. A collection on GitHub now lists numerous Jev alternatives, and a public comparison ranking called JevBench includes more than 100 open models.

Technically, many of them rely on the same principle: they take an existing open language model, often from the Qwen family of the Chinese company Alibaba, and retrain it so that it no longer outputs text but only probabilities for the given answers. This makes them fast and cheap, because each decision only needs a single computation pass.

What Jev is and how it is being used is covered in separate articles. This one is about the competition.

The big providers: OpenAI, Microsoft, Cloudflare and Perplexity

OpenAI: its Decisions interface launched in early October, powered by the small model GPT-6 Luna. It handles yes-or-no questions, choices and ratings, including for images. According to Forbes, it is up to ten times faster than GPT-6 Luna in normal text mode and costs 10 US cents per million input tokens, with outputs free. That is more expensive than Jev, but it is part of the familiar OpenAI offering.

Microsoft: Microsoft-Decision-1 appeared on 9 October via Microsoft Foundry and OpenRouter. It is based on Qwen3.5-9B and, at 4.2 US cents per million tokens, costs exactly as much as Jev. Microsoft says it was the most accurate in an internal evaluation with almost 150,000 questions and 35 times faster than GPT-6 Sol. This has not been independently verified.

Cloudflare: the Clef and Clef Flash models appeared on 1 October as open models that also run on Cloudflare's network. On 9 October came Clef-omni, which can assess audio and video as well as text and images, for example to screen content in advance.

Perplexity: Decider V1.1 has 27 billion parameters, can be downloaded freely under the Apache 2.0 licence and costs only 2 US cents per million tokens through Perplexity's interface. Perplexity says it reaches 94.5 percent accuracy on the Decision Bench test at 1.7 US cents per 1,000 decisions.

There is also Liquid AI with D1 for 4 US cents, which additionally comes as small open variants for edge devices, Inception's Mercury Decide, which is also available for free on OpenRouter with rate limits, and Upstage with Solar Decide.

The open models: Amazon Strands Decider, H2O-Lightning and co.

Models that can be downloaded and run in-house are particularly interesting for companies with sensitive data. The data then never leaves their own data centre.

Amazon Strands Decider 2B: on 1 October, Amazon Web Services released a small open model with around two billion parameters, built on Qwen3.5-2B. It is under the Apache 2.0 licence, together with training data and scripts. According to Amazon, it solves all easy tasks of the JevBench comparison ranking correctly, meaning typical routine decisions in AI agents. VentureBeat called it a free, fast Jev killer whose strength lies less in record scores than in being open and self-hostable. Amazon also offers a managed service through its Bedrock cloud platform.

H2O-Lightning-4B: the company H2O.ai reports that its open model with four billion parameters ranks first out of 104 open models in the JevBench overall score, with 72.5 out of 100 points. Jev itself, which runs as a reference, scores 71.5. The score weights accuracy, reliability of probabilities, speed and cost equally. According to H2O.ai, it costs an estimated 2.1 US cents per 1,000 decisions. These figures come from a manufacturer press release, and only about one point separates the top.

There are also projects by independent developers such as Jared Palmer's Kev, available in sizes from 0.8 to 27 billion parameters, and Laya from Convai Innovations, which covers more than 100 languages with very small models. Together AI also publishes model weights and a training recipe with Tev1.

Important: open does not mean free. Running a model yourself requires computing power, specialists and a careful check of whether the model's probabilities are accurate.

A common language: the System One interface

A quiet but important development concerns the way programs talk to these models. According to The New Stack, TypeSafe's format for requests and responses is becoming a common standard. Upstage offers Solar Decide through this System One format, Amazon's Strands Decider is designed for it, and the popular program Ollama, which many people use to run AI models on their own computers, added a matching interface in version 0.35. OpenAI, by contrast, is going its own way with its Decisions interface.

For developers, this means: if you write your software once for this format, you can later switch models more easily, for example from Jev to an open model on your own servers. That reduces dependence on a single provider.

The New Stack also points out a limit: a common format does not guarantee that the models answer equally reliably. A model's probabilities have to match the actual outcome. If a model states 80 percent confidence, it should be right in about 80 percent of cases. H2O.ai names exactly this property, calibration, as the value its team watches most.

Who is ahead and what this means for Europe

In usage, Jev is far ahead. Through the OpenRouter platform, around 1.08 billion requests went to Jev in the week to 14 October, while second-placed Perplexity Decider reached 43.2 million. These figures only show part of the picture, though: anyone running an open model themselves does not appear in any statistics.

In quality, by contrast, the field has moved close together. Forbes stresses that in the end, rankings matter less than how often a model is right on your own data and what a mistake costs. A model that costs half as much but sends twice as many cases to people for review saves little.

For companies in Germany and the EU, the competition is good news. According to TypeSafe, Jev processes data in the US. With open models such as Strands Decider, H2O-Lightning, Kev or the open variants from Cloudflare, Perplexity and Liquid AI, decisions can be made on your own servers in Europe without passing data on. For developers and hobbyists: such models can now also be tried on your own computer via Ollama.

Our assessment: the speed with which an entire market is emerging here is remarkable. But it also shows that the underlying technology is comparatively easy to replicate. Jev's lead currently lies mainly in adoption and name recognition, less in an unassailable technical advantage.