What Are Decision Models? The AI That Can't Write a Sentence
TypeSafe, OpenAI, Microsoft and Liquid AI all shipped one in about a month. Strip the buzzword and it's a classifier: it picks an answer and never writes a word. Here's where you already meet one, and where it can go wrong.
- 'Decision model' is the industry's new name for a classifier. It picks from answers you give it, says yes or no with a confidence number, or gives a score. It never writes text.
- Spam filters were classifiers too. The new part: you describe the job in plain English instead of training it on thousands of labeled examples.
- It's cheap. At TypeSafe's posted $0.042 per million input tokens, sorting 10,000 short emails works out to about 21 cents.
- An independent trolley-problem test found Jev, TypeSafe's classifier, changed its answer in 370 of 2,620 paired questions when only the order of the options was swapped. Keep a human on anything that matters.
A 'decision model' is the AI industry's term for a classifier. Instead of writing text like ChatGPT, it picks one answer from options you supply, returns a yes/no with a confidence number, or gives a score. TypeSafe's Jev, OpenAI's Decisions API, Microsoft-Decision-1 and Liquid AI's d1 all shipped between September and October 2026.
In about a month, four companies shipped the same kind of AI. TypeSafe launched Jev, a classifier, on September 15. OpenAI put out a Decisions API. Liquid AI released d1 as a free download on October 7. Microsoft followed with Microsoft-Decision-1 on October 9.
The industry calls this new category “decision models.” Some vendors call it “System One.” Strip the branding and you get something older and simpler: a classifier. It picks an answer from options you give it, gives a yes or no with a confidence number, or hands back a score. It never writes a sentence.
So why is the AI that can’t talk suddenly everywhere? And where is one already making calls about you?
What is a decision model, in plain English?
Think of a mailroom clerk with three bins. Every envelope goes in one bin. The clerk doesn’t write you a memo about the envelope. They sort it.
That’s the whole job. You hand it some text and a question with fixed answers. “Is this email spam?” Yes, 97% sure. “Which team handles this ticket: billing, shipping or tech?” Shipping. “Rate this review from 1 to 5 for how angry the customer is.” A 4.
Your spam filter has been a classifier for years. So has the fraud alert on your credit card. None of that is new.
Here’s what changed. The old ones needed thousands of hand-labeled examples and a data scientist to train them. The new ones read a plain-English description of the job, like instructions you’d type into ChatGPT, and get to work. I broke down the leading one in Jev, explained.
Why is every AI company launching one right now?
Because they’re cheap, fast, and a lot of AI work inside apps was never writing. It was sorting, billed at writing prices.
Vercel runs a gateway developers use to switch between AI models. It says Jev reached nearly 13% of its paid teams within 24 hours: about twice the share of OpenAI’s GPT-5.6 family and more than six times Anthropic’s Fable 5.1.
Vercel called it the fastest adoption in its gateway’s history, then added the honest part: the real test is whether people stick around.
TypeSafe raised a $40 million seed round led by DCVC the day Jev launched. Its posted price is $42 per billion input tokens (tokens are the word chunks AI companies bill by). That’s $0.042 per million, and output isn’t metered at all. TypeSafe claims answers come back in 70 to 500 milliseconds.
What does that mean for a real business? Here’s the math:
- Say each email plus your sorting instructions is about 500 tokens.
- 10,000 emails × 500 tokens = 5,000,000 tokens.
- 5 million ÷ 1 million × $0.042 = $0.21.
Twenty-one cents to sort 10,000 emails.
Two catches. TypeSafe hasn’t published results on public benchmarks and tells users to run their own tests. And its own site asks whether the prices are temporary or subsidized without answering.
Microsoft’s version, built on a retooled Alibaba Qwen model, starts at the same $0.042 per million. Liquid’s d1 is free to download and run on your own hardware; see my Liquid AI d1-3B verdict.
SourcesVercel: Jev is the fastest-adopted model in AI Gateway history, The Next Web on Jev's launch, funding and pricing, TestingCatalog on Microsoft-Decision-1.
Where do you already run into a classifier?
Here’s where they show up, and what a wrong call costs you.
| Where you meet it | The question it answers | What a wrong call costs you | What to do |
|---|---|---|---|
| Email spam filter | Spam or not? | A real invoice or job offer lands in junk | Check your spam folder weekly |
| Card fraud alert | Fraud or normal purchase? | A blocked card at the register | Answer the bank’s text, ask for a review |
| Social media moderation | Allowed, limited or removed? | Your post disappears, or your account gets flagged | Use the appeal button, ask for a person |
| Customer support routing | Billing, shipping, tech or urgent? | Your ticket sits in the wrong line for days | Reply with clear words like “refund” or “cancel” |
| Loan or rental pre-screen | Pass to a human, or decline? | You get filtered out before anyone reads your file | Ask for a manual review |
They’re fine when a mistake is small and fixable. They get scary when nobody checks the bin.
Can a small change in the question flip the answer?
Yes, and somebody measured it. Zach Landes ran the trolley problem (save five people by diverting a train into one?) through 14 classifiers: Jev, OpenAI’s GPT-6 Luna through its Decisions API, eleven local models and one untuned baseline. 5,240 setups in all.
The test that matters for you: the same question, with the two tracks described in the opposite order. Nothing else changed.
Jev flipped in 370 of 2,620 pairs. That’s 14.1%. GPT-6 Luna flipped in 330. Asked the identical setup three times, Jev’s answers landed on different sides of yes/no in 361 of 5,240 setups.
It gets stranger. In 480 matched pairs, Landes swapped the person on the side track for Jev’s own server. In 180, Jev would sacrifice a stranger to save the crowd but not its server. GPT-6 Luna was one of only two models that never put its server ahead of people.
His bottom line is blunt:
“No one should be asking an AI to rule on life or death.”
Now swap trolley tracks for your loan file, your moderation appeal, your fraud alert. If listing the options in a different order can move the answer, it’s fair to ask what the order of fields on a form does.
Should a small business actually use one?
For sorting, yes. Tagging inbound leads as hot, cold or spam. Routing support tickets to the right person. Flagging one-star reviews for you to read first. Those are high-volume, low-stakes calls, and 21 cents per 10,000 is hard to argue with.
Do you need to code? Today, mostly yes. These are developer tools, so most owners will meet one inside software they already pay for, or through whoever builds their site. Start with my small business AI stack.
Two rules if you do. Test it on a few hundred of your own real messages before you trust it, because TypeSafe hasn’t published public benchmark results. And never let it make the final call on a refund, a firing, a loan or an account ban. Let it sort the pile. A person decides.
So what’s the call?
“Decision model” is a new label on an old, useful idea. The classifier got easier to set up and absurdly cheap, and that’s real.
It didn’t get wiser. It can’t explain itself, and Landes showed the answer can move when the question gets reshuffled. Use it as the mailroom clerk. Never as the judge.
Micah Berkley, founder and editor of Run the Eval, where the call on AI comes from evidence, not hype. #TheAIMogul
Bottom lineDecision models are a real upgrade for boring, high-volume sorting: spam, support tickets, leads, reviews. They are not judges. Use them to sort the pile, and keep a person on any call that touches someone's money, job, health or account.