Summarize this article with:
DeepSeek V4.1 Flash and Jev, the fastest-adopted model on Vercel's AI Gateway, point to the same shift: AI decisions for a fraction of a cent per thousand calls.
Two launches in the last few weeks say a lot about where AI spending is going, and neither is a bigger model.
The first is DeepSeek V4.1 Flash, a fast, cheap model that DeepSeek now uses to serve requests sent to V4 Pro, billed at Flash rates, according to TechNode. The second is Jev, a model from TypeSafe that cannot write a single word. It reached 13% of Vercel AI Gateway's paid teams within 24 hours of launch, the fastest adoption in that gateway's history, as The Next Web reported.
Together they point to the same shift. A growing share of AI calls in production are not asking for an essay. They are asking for a decision: which team, which category, is this a refund, how urgent is it. And those decisions are getting cheap enough to run on every request.
What a decision costs now
We sent the same support ticket, "I was charged twice for my October invoice and I need a refund today", to route to billing, technical or sales, through Eden AI's API on 21 October 2026. Three setups, all answering correctly:
| Setup | What comes back | Median latency | Cost per 1,000 decisions |
|---|---|---|---|
| Jev, decision model | "billing", with a probability for each team | 318 ms | $0.015 |
| DeepSeek V4.1 Flash, route A | "billing", as text | 868 ms | $0.0055 |
| DeepSeek V4.1 Flash, route B | "billing", as text, after reasoning first | 2,064 ms | $0.038 |
At these prices, cost stops being the reason not to put a model in front of every request. A thousand decisions cost between half a cent and four cents. A million cost a few dollars to a few tens of dollars.
The table also hides a lesson. The same model, DeepSeek V4.1 Flash, cost seven times more on route B than on route A, because that route reasoned before answering and spent around 70 output tokens to write one word. With text generation, the price per token is only half of the bill. The other half is how many tokens the model decides to produce.
From generating to deciding
A decision model works differently from a chat model. You send a state, the text you want judged, and typed questions about it. The model reads everything once and returns a value for each question, without writing any text:
- Yes or no: a probability that something is true, such as "is this a refund request?"
- One option from a set: the chosen option plus the probability of every option, such as "which team should handle this?"
- A level on a scale: a score on your own rubric, such as "how urgent is it?"
That design has three consequences. The answer is always one of the options you defined, so there is nothing to parse and no invented category. Every answer carries its own confidence, so your code can act when the model is sure and escalate to a person when it is not. And since nothing is generated, there are no output tokens to pay for: Jev is billed on input only, at $0.042 per million tokens.
Jev is not alone. Through Eden AI's decisions endpoint today you can also call GreenPT's Metis, Microsoft's decision model on Azure, a decision model from OpenAI and a preview from Qwen. When several labs ship the same new category within weeks, it is usually a sign that the need was already there.
Which tool for which job
| Task | Best fit | Why |
|---|---|---|
| Routing a ticket, a lead or a request | Decision model | Fixed options, a confidence score, no parsing |
| Gating: is this spam, a refund, ready to send? | Decision model | A probability your code can threshold |
| Pre-filter before an expensive model | Decision model or a cheap fast LLM | Costs a fraction of the call it can avoid |
| Short extraction or a one-line answer | Cheap fast LLM | The output is words, not a fixed option |
| Drafting, explaining, coding | General LLM | Decision models cannot write text at all |
The most useful pattern combines them. A decision model reads every incoming request and decides what happens next: answer automatically, send to a cheap model, send to a frontier model, or hand to a person. The expensive model only runs when it is actually needed.
What to keep in mind
- Vendor numbers are vendor numbers: TypeSafe reports Jev up to 193.6x faster and 444.6x cheaper than LLMs on its workflow evaluations. Your ratio depends on your prompts, your current model and your traffic. Our own test above shows a smaller, still meaningful gap.
- Test on your own data: a decision model is only as good as the questions and options you give it. Phrase questions positively and keep options distinct: negations are easy to get wrong.
- Check the confidence, not just the answer: the probabilities are the point. Decide in advance which threshold triggers automation and which one triggers a human review.
- Watch output tokens on cheap LLMs: a model that reasons before answering can multiply the cost of a one-word reply, as route B shows. Check
usageon real requests, not just the price list. - Check the region: Jev is served from the US today. If your data must stay in Europe, Metis and Microsoft's decision model have EU routes on Eden AI.
Try both through Eden AI
A decision through Eden AI's decisions endpoint:
import requests
response = requests.post(
"https://api.edenai.run/v3/alpha/decisions",
headers={"Authorization": "Bearer YOUR_EDENAI_API_KEY"},
json={
"model": "typesafe/jev-latest",
"state": "Hi, I was charged twice for my October invoice and I need a refund today.",
"questions": {
"is_refund": {"type": "noul", "instructions": "Is this a refund request?"},
"team": {
"type": "choice",
"instructions": "Which team should handle this?",
"criteria": {
"billing": "Payments, invoices, refunds",
"technical": "Bugs, outages, integrations",
"sales": "Pricing, upgrades, new accounts",
},
},
},
},
)
print(response.json()["answers"])Swap the model string for greenpt/metis or azure/microsoft-decision-1@eu to compare decision models on the same questions. The decisions endpoint is in alpha, so its request and response shapes may still change; the decisions documentation has the details.
The same ticket through a cheap LLM uses the standard chat completions endpoint, with the model set to a DeepSeek V4.1 Flash route of your choice. Compare latency, cost and the usage field on your own requests before picking one.
Where to start
Pick one decision your product makes thousands of times a day, a ticket route, a spam check or a lead score, and run it through a decision model next to what you use today for a week. If the answers hold and the confidence scores behave, you have moved a recurring cost from cents to fractions of a cent, and freed your larger models for the work that needs them. For more on matching model cost to task, see our guide to price-performance routing.
Frequently asked questions
What is a decision model?
A decision model answers typed questions about a piece of text, such as yes or no, one option from a set, or a level on a scale, and returns a value with its probability instead of generating text. It reads the input once, which makes it fast, and it is usually billed on input tokens only.
How cheap are AI decisions now?
In our test on 21 October 2026, routing one support ticket cost between $0.0055 and $0.038 per thousand decisions, depending on the model and route. That is a fraction of a cent per thousand calls for a cheap LLM, and about a cent and a half per thousand for Jev.
Why is Jev being adopted so fast?
Jev reached 13% of Vercel AI Gateway's paid teams within 24 hours of launch. It targets a common need, small repeated decisions inside applications, with typed answers, confidence scores, a low price and no output tokens to pay for.
Should I replace my LLM with a decision model?
Only for decisions. A decision model cannot write, summarise, extract free text or explain itself. Use it for routing, classification, gating and scoring, and keep a general model for anything that needs words.
Can I use decision models in the EU?
Yes, with the right model. On Eden AI, GreenPT's Metis and Microsoft's decision model have EU routes. Jev is served from the US today.
