Decision models demo

Decision model vs regular LLM

Same question, two kinds of model, both called live. The decision model returns a probability for every option. The regular LLM writes JSON that your code has to parse. Pick a use case and race them.

Decision model Microsoft-Decision-1

0.00 s

Correct
–
Median per call
–
Tokens in / out
–
Probabilities
for every option
Cost this run–

Regular LLM GPT-6 Luna

0.00 s

Correct
–
Median per call
–
Tokens in / out
–
Probabilities
none, a label only
Cost this run–

Example JSON: request and response for one item

Decision model · request


          
Decision model · response

          
Regular LLM · request

          
Regular LLM · response

        

      

    

  


  

    
All four use cases

    
Median model latency per call, measured on the server, and how many answers were right. 

    

    
Decision modelRegular LLM

  


  

    
What it costs

    
Cost per decision, from the tokens each model really used, at the published Global pay-as-you-go prices (dated below). 

    

    
Decision modelRegular LLM

    

      

        
ModelPer callPer runPer 1,000 decisionsPer 1M decisions

        
      

    

    

  


  

    
How it works

    
    
      
  1. Your code asks typed questions: pick one option, place it on a scale, yes or no.
    2. 
      
  2. The decision model answers with a probability for every option. The LLM gets the same questions in a prompt and writes JSON.
    3. 
      
  3. Your code decides what happens: thresholds, policies, escalation when confidence is low. The model never runs free.
    4. 
    

    
Both models run on Microsoft Foundry. Latency is the model round trip measured on the server; the big clocks are wall time in your browser. The LLM gets a clear JSON answer format, runs with reasoning effort none (its fastest setting) and at most 300 output tokens per call. Limits: 10 live runs per minute per visitor, 500 model calls per day.