Decision models, explained: typed answers instead of JSON you have to parse
Jev, Perplexity Decider 27B, Microsoft-Decision-1 and Gemini Decision 1 return probabilities, not prose. How they work, what they cost and which one to pick.
Most of what production apps ask an LLM isn't "write me something". It's a decision. Is this ticket urgent? Which team should get it? How angry is this customer? Is this image safe to publish?
The usual way to get those answers is to ask a chat model to "reply with JSON", then validate the JSON, retry when it's malformed, and hope the model didn't invent a category you never offered. Decision models remove that whole layer.
What a decision model does
A decision model never writes text. You give it two things:
- The state: the thing to decide about. A support ticket, a transcript, a JSON record, and for some models an image, audio, video or PDF.
- Typed questions about that state.
It returns exactly one typed answer per question, each with a calibrated probability, in a single fast pass. It can't answer outside the options you gave it, so there's nothing to validate.
There are three question types:
| Type | Returns | Example |
|---|---|---|
noul | A yes/no probability | "Does this need a reply within the hour?" → 0.88 |
choice | One label from criteria you describe | "billing", "technical" or "other" → billing, 0.97 |
score | A position on a rubric you define | "No signal" → "Threatens to cancel" → 1.8, 0.81 |
One request, three answers
That's three decisions for well under a thousandth of a cent. Because you get probabilities, you choose the thresholds: auto-route above 0.9, send to a human between 0.6 and 0.9, ignore below.
The four models
All four take the exact same request on POST /v1/decisions. Change the model field and run your own data through each.
[Image blocked: Jev decision model on CompanyFabric]
Jev [blocked] by TypeSafe is the default when you omit model. Typed decisions in about 100 ms. $0.042 per 1M input tokens, output free. Model id typesafe/jev-latest.
[Image blocked: Perplexity Decider 27B on CompanyFabric]
Perplexity Decider 27B [blocked] works the same way and can also read images in the state: screenshots, receipts, product photos. $0.04 per 1M input tokens, output free. Model id perplexity/decider-27b.
[Image blocked: Microsoft-Decision-1 on CompanyFabric]
Microsoft-Decision-1 [blocked] launched in Microsoft Foundry in October. In Microsoft's own 36-benchmark comparison it ranked first on accuracy and was the fastest model measured. $0.042 per 1M input tokens, output free. Model id microsoft/microsoft-decision-1.
[Image blocked: Gemini Decision 1 on CompanyFabric]
Gemini Decision 1 [blocked] is Google's decision model and the most multimodal of the four: the state can be text, JSON, an image, audio, video or a PDF. $0.10 per 1M input tokens, output free. Model id google/gemini-decision-1.
Which one to pick
- Text only, high volume: start with Jev. It's the default for a reason.
- Screenshots and photos: Perplexity Decider 27B.
- Calls, meetings, videos, contracts as PDFs: Gemini Decision 1.
- Not sure: run the same 200 labelled examples through all four and keep the one that agrees with your labels most often. It's one field to change, and the whole comparison costs cents.
Where decision models fit
Ticket triage, lead scoring, content moderation, routing an agent's next step and grading evals. Anywhere you currently ask a chat model to "answer with JSON" and then hope it does.
The full reference is in the Decisions API docs [blocked].
- Decision models
- Jev
- Perplexity
- Microsoft
- Gemini
- Classification
Every model in this post is behind one key and one balance. Free starter credits on signup.
Get API KeyKeep reading
Claude Code with any model: GPT, Gemini, DeepSeek and Kimi on one key
Set two environment variables and Claude Code runs on any model in the CompanyFabric library, with prompt caching intact on Claude and a budget cap on every key.
New on CompanyFabric: Claude Fable 5.2, Gemini 4 Argon, Muse Wonder, MiMo V2.6, Kimi K3.1 and decision models
Five frontier model families from Anthropic, Google, Meta, Xiaomi and Moonshot, plus decision models from TypeSafe, Perplexity and Microsoft. One key for all of them.
Kimi K3.1 Pro in Europe: run Moonshot's flagship on Verda GPUs in Germany
Bring your own Verda key and Kimi K3.1 Pro runs on GPUs in Germany instead of Moonshot's servers in China, at a 0% fee. Same model id, same code.