Jev and decision models: what they change on the phone
Decision models are AI models that don't write text: they take a question and a list of possible answers, pick one and say how sure they are, in a fraction of a second. The first to launch one was Jev by TypeSafe AI, introduced in mid-September 2026. October brought Amazon's Strands Decider and OpenAI's Decisions API, which since 15 October plugs straight into OpenAI's voice model. For any business that lets an AI answer the phone, this matters: the moment when the assistant decides what to do with a call can become almost instant.
This article covers what was released, what the numbers say (all of them published by the vendors) and what it may change for an AI phone assistant.
What decision models are
A model like ChatGPT or Gemini generates sentences. That is great for conversation, but slow when all you need is a short answer, and hard to control. Ask it "is this call urgent?" and you may get "yes", "probably" or a whole paragraph.
A decision model works the other way round. You give it the question and the shape of the answer (say: "urgent", "call back", "booking", "information") and it returns exactly one of those options, with a probability for each. TypeSafe calls them "System One models": models built for fast, structured decisions that software can use directly.
Who launched them, and the numbers
| Model | Who | When | What the vendor says |
|---|---|---|---|
| Jev | TypeSafe AI | mid-September 2026, early access | 70 to 500 milliseconds per decision; up to 255 options per question; very low input price and free output (announcement) |
| Strands Decider 2B | Amazon (Strands Labs) | 1 October 2026, open source | a small 2-billion-parameter model, light enough to run locally (TechCrunch) |
| Decisions API (gpt-6-luna) | OpenAI | beta on 6 October, generally available on 22 October | answers "about 10x faster" than the Responses API; up to 400 questions per request; native GPT-Live integration (official guide, changelog) |
Two caveats. First, none of these numbers has been checked independently yet: TypeSafe itself says its comparisons rely on its own evaluations and that the reference may favour some models. Second, the Jev announcement says nothing about languages.
According to TechCrunch, "dozens" of similar models appeared after Jev. That is what has driven the buzz of the last few weeks: not one product, but a new kind of model.
What changes for a call handled by AI
On a call, the assistant doesn't just have to sound good. While the caller is talking, it has to make a string of small decisions:
- what the caller wants: book, move an appointment, ask about opening hours, speak to someone;
- whether it is urgent: at a medical practice, severe pain is not a request like any other;
- whether to hand the call to a person: and to whom, under the business's forwarding rules;
- whether it has understood enough: or needs one more question before acting.
Today these decisions are usually made by the same model that holds the conversation, and every step adds waiting time. A decision model can make them in parallel, in under half a second according to the vendors, with an answer the software reads without guesswork. OpenAI names this use case explicitly: GPT-Live can "call Decisions directly", or decisions can be triggered as the transcript of what the user says comes in.
For the caller that means three things: fewer silences while the assistant "thinks", more consistent routing (the same request always ends up in the same place) and a faster handover to a person when one is needed.
Who benefits most
- Medical and dental practices: spotting urgent calls and passing them to the practice is the most sensitive decision. A clear, fast rule ("urgent: yes or no") is easier to check than a written answer. See how routing works today for medical practices and dental practices.
- Hotels: a new booking, a guest already on site with a problem, a simple question. Three different paths, to be chosen in the first few seconds (hotels).
- Real estate agencies: working out whether the caller wants to sell, buy or rent, and which property they mean, decides who calls them back.
What to do now
Nothing urgent: these models are tools for the people who build assistants, not something to switch on. If you use or are considering an AI phone assistant, they do suggest good questions for your provider:
- how calls are routed, and how fast;
- under which rules a call goes to a person, and whether you can change them;
- what happens when the assistant isn't sure it has understood.
To hear how an AI phone assistant answers your business's calls today, you can try the demo.
In short
- Decision models pick an answer from a list instead of writing: they are faster and more predictable.
- Jev by TypeSafe AI led the way in September; in October came Amazon's Strands Decider and OpenAI's Decisions API, wired into GPT-Live.
- On the phone they are useful to route calls, spot urgent ones and decide when to hand over to a person.
- All the numbers still come from the vendors: wait for independent tests and real calls in your language.
In the same weeks OpenAI also released GPT-6.1 Terra (26 October), a text-and-image model without audio: for real-time calls it matters less than voice models such as Grok Voice 3.0.
Frequently asked questions
What is Jev by TypeSafe AI?
It is an AI model that TypeSafe AI introduced in mid-September 2026 as the first "System One model". Instead of writing text, it returns a structured decision (a choice from a list, a score, a yes or no) with a probability. It is in early access, with a waiting list.
How is a decision model different from a chatbot?
A chatbot generates sentences one word at a time. A decision model only answers within the options you give it and tells you how confident it is. It is faster and more predictable, but it can't hold a conversation: it works alongside a model that talks, not instead of it.
Does OpenAI's Decisions API work with voice?
Yes. Since 15 October 2026 GPT-Live, OpenAI's voice model, can call the Decisions API directly, or decisions can be triggered as the transcript of what the user says comes in. The Decisions API has been generally available since 22 October.
Does a small business need to do anything?
No. These are tools for the people who build assistants, not for the businesses that use them. What a business can do is ask its provider how calls are routed, how fast, and under which rules a call is passed to a person.
