Decision models, the useful 20%

The short version. A decision model is an AI that cannot write a sentence. You give it some text, a question, and the answers it is allowed to give. It hands back one of your answers, usually with a number saying how sure it is. It does that in a tenth of a second for a fraction of a cent. Three big names landed in three weeks: TypeSafe's Jev (Sep 15), OpenAI's Decisions API (Sep 29), and AWS's Strands Decider 2B (Oct 1). All three names in your Brief check out.

The one idea: a multiple-choice test, not an essay

A normal chatbot (an LLM, or large language model) writes its answer word by word. It can say anything, including things that aren't true. A decision model only fills in the bubble. You write the question and the options; it limits the possible answers before it starts, so it can't wander off and can't invent an option you didn't offer.

The tool everyone copied is Jev. TypeSafe calls it a "System One" model, borrowing psychologist Daniel Kahneman's term for fast, gut-level thinking. Three question types cover almost everything:

Yes or no. "Does this email need a reply today?" You get back a probability, like 0.88.
Pick one. "Which team should get this ticket: billing, shipping or tech?" You get a share of confidence for each option.
Score. "How urgent is this, from cosmetic to blocking?" You get a position on the scale you wrote.

TypeSafe (and the open copies) call the yes/no type a "noul". It just means a yes/no with a probability attached.

Where it sits in an agent

An agent (an AI that takes actions, not just answers) makes dozens of small choices per job: which tool next, is this safe, is it done? Today a big model often makes each of those calls, slowly and expensively. The new pattern, as the AI Daily Brief laid it out, splits the work:

1 · UnderstandA language model reads the messy request.

2 · DecideA decision model picks the next step from a fixed list.

3 · ActOrdinary tools and code do the thing.

4 · CheckMonitoring watches what happened.

How this differs from an agent framework. A framework (Strands Agents, LangGraph) is the plumbing: the loop that calls models and runs tools. A decision model is one small part you plug in at the forks in that pipe. Telling detail: Strands Decider comes from the team behind the Strands Agents framework, built to sit inside it for routing, tool picking and guardrails.

How it differs from a plain LLM call. You could ask a chatbot "answer only billing, shipping or tech." But it still writes text, which you then have to read and check. A decision model skips the writing. Jev claims 70 to 500 milliseconds a call and charges $0.042 per million tokens of input, with output free. (Tokens are word pieces; a million is a few novels.)

The three, side by side

Tap a column heading to sort. Tap a filter to narrow.

Product	Maker	Launched	Runs where	You get back	Status
Jev	TypeSafe (founder Diogo Almeida)	Sep 15, 2026	TypeSafe's cloud, paid per use	Your answer plus a calibrated probability for every option	Generally available
Decisions API	OpenAI, on GPT-6 Luna	Sep 29, 2026	OpenAI's cloud	One of your answers; text or image input	Limited preview, no public docs yet
Strands Decider 2B	AWS Strands Labs	Oct 1, 2026	Your laptop or server (about 2B parameters)	Choice, yes/no or score, each with a confidence	Free, Apache-2.0, training recipe published
Laya (related)	Convai Innovations	Days after Jev	Your machine (421M parameters, English)	Same three question types as Jev	Free, Apache-2.0

Laya's exact release day isn't stated on its model page; it sorts just after Jev. "Parameters" are the model's adjustable knobs, a rough size measure.

Each one, opened up

Jev · TypeSafe · the original Name verified

What it is. TypeSafe's first model, introduced on September 15, 2026 by founder Diogo Almeida. Its pitch: "unstructured state in, typed probabilistic decisions out." In plain words, messy text goes in, a labelled answer with odds comes out.

The trick. TypeSafe trains it with what it calls RLCD, Reinforcement Learning for Calibrated Decisions. "Calibrated" means honest odds: of everything it calls 80% likely, about 80% should turn out true. That's what lets you write a rule like "below 0.8, ask a human."

Uptake was fast. VentureBeat reported about 13% of Vercel's paid AI Gateway teams ran it within a day, and Cloudflare, LangChain and Langfuse added it within three days.

Spelling note. The company writes itself "TypeSafe" (sometimes "TypeSafe AI"). "TypeSafe Jev" in your Brief is right. You already have a hands-on Jev primer from September 22.

Decisions API · OpenAI · the big-platform version Real, thinly documented

What it is. Announced at OpenAI DevDay on September 29, 2026. OpenAI's own line: it "uses Luna to classify inputs, route requests, or choose an action from predefined answers." Luna is OpenAI's small, fast GPT-6 model. The OpenAI Developers post sells it as "real-time decision-making."

How it differs from Jev. It accepts images as well as text, and returns one chosen answer; Jev also returns odds for every option. Reported speed is about 150 milliseconds.

What's missing. It's in limited preview for selected customers. As of September 30 there was no public endpoint reference, schema or price. Treat any detailed how-to you find online as guesswork until OpenAI publishes docs.

Strands Decider 2B · AWS · the open one you can run at home Name slightly off

What it is. Released October 1, 2026 by AWS's Strands Labs (authors Marc Brooker, Mike Chambers and Fabio Nonato de Paula). It's a small add-on to an existing open model, Qwen3.5-2B, with a tiny scoring head bolted on in place of the part that writes text.

Speed. A median of about 115 ms on a gaming graphics card and ~153 ms on an M3 MacBook. That matches your Brief's "under 150ms" on the GPU; on a Mac it's a hair over.

Open in full. Weights, code and training data are public under Apache-2.0 (free to use, even commercially), on GitHub. It scores 72.3% on JevBench, a test set built around Jev-style questions. AWS is plain about the limits: it's "significantly worse at solving complex problems than reasoning models."

Name check. The official name is Strands Decider 2B, from AWS. "Amazon Strands Decider" is fine shorthand, but no product goes by exactly that. Coverage: Tech Times, MarkTechPost.

Laya · Convai Innovations · the tiny open copy Related

One of many open models that copied Jev's question format. About 33 milliseconds a question, built on a much smaller model (421 million parameters), with yes/no, pick-one and score questions, Apache-2.0. Worth knowing because it shows the format, not any one company, is what's spreading.

The catch: it can still be talked into the wrong answer

"It can't hallucinate" is true. "It can't be fooled" is not. A decision model reads whatever text you hand it, and an attacker can plant text there. That's prompt injection (hidden instructions or fake evidence slipped into the input).

Check Point (a security firm) broke Jev's verdict in all nine attack setups it tried, for about 50 cents a break, by slipping fake audit opinions into documents. Their lesson: "Check what goes in, not just what comes out."
A research paper on arXiv found injected text shifts Jev's odds but rarely flips it to the attacker's choice; success rose from 1.8% to 3.5% with smarter attacks. Smaller than with chatbots, not zero.
VentureBeat showed one test where a "block this" probability fell from 0.76 to 0.48 after injected text, and noted that even reordering the options can move the answer.

The AI Daily Brief put it simply: every new model in the chain is a new place for things to fail.

Which one fits this job?

Tap a job to see the pick.

Try it on your own Mac, free, private data stays home.

Strands Decider 2B. Runs locally, no bill, open recipe. Laya if you want it even smaller and faster.

Sort or route real traffic now, with odds you can set a threshold on.

Jev. The only one of the three that's generally available with per-option probabilities. Check its odds against your own data before trusting them.

Your app already runs on OpenAI, or the input is a picture.

Decisions API, once you get preview access. Until there are public docs and prices, don't plan around it.

Someone (a customer, a regulator) can demand the reasoning.

None of them. They return a pick and a number, no argument. Use a full LLM or a person. AWS says its model is weak at complex problems.

The text comes from emails, web pages or uploads you don't control.

Any, with guards. Screen the input, keep a human approval on anything costly, and log every decision, as VentureBeat and Check Point advise.

The Keys thread. Your Brief filed this as a fresh instance of "who decides what happens next." The useful nuance: these models only choose from a menu a person wrote. So part of the deciding moves to whoever writes the menu and sets the threshold, and part moves to a fast model nobody watches call by call.

Words you'll see

Decision model

An AI that picks from answers you supply instead of writing its own.

System One

Kahneman's fast, gut-level thinking. TypeSafe's name for the category.

Calibrated

Honest odds. True across many answers, never a promise about one.

Routing

Deciding which tool, team or model gets the next piece of work.

Guardrail

A quick "is this safe?" check before an agent acts.

Prompt injection

Planted text in the input that steers the model's answer.

Deliberately not here

How to code against each API (see the Jev primer for a working call). Benchmark tables, which age in weeks. The dozens of smaller open copies beyond Laya. And how the models work inside, which changes nothing about how you'd use them.

Sources

TypeSafe, Introducing System One models and Jev (Diogo Almeida), Sep 15, 2026
VentureBeat, Companies are putting Jev in charge of AI agent decisions (Louis Columbus), Sep 21, 2026
Wu & Lim, Decision Hijacking: Prompt Injection Attacks on Jev's Typed Probabilistic Decisions, arXiv, Sep 23, 2026
Check Point, A Decision Model Breaks Like Any Other Language Model, Sep 24, 2026
OpenAI Developer Community, DevDay 2026 announcements and developer resources, Sep 29, 2026
OpenAI Developers on X, Decisions API, powered by GPT-6 Luna, Sep 29, 2026
Firecrawl, OpenAI's Decisions API vs Jev (Hiba Fathima), Sep 30, 2026
Strands Agents, Introducing Strands Decider 2B (Brooker, Chambers, Nonato de Paula), Oct 1, 2026
Hugging Face, StrandsAgents/strands-decider-2B model card, Oct 2026
GitHub, strands-labs/strands-decider, Oct 2026
MarkTechPost, AWS Strands Labs Releases Strands Decider 2B, Oct 1, 2026
Tech Times, AWS Releases Decision Model for AI Agents, Oct 2, 2026
AI Daily Brief, Decision models are becoming a new AI category, and a new layer of risk, Oct 2, 2026
Hugging Face, convaiinnovations/laya model card, accessed Oct 4, 2026

Rich Price · primer · October 4, 2026 · Decision models, the useful 20%