How Does CompletionPrism™ Choose a Model for Each Prompt, Step by Step?

CompletionPrism™ scores the prompt, sizes the work, prices every allowed model from graded study runs, picks the lowest expected cost, and shows the actual spen

By Avik Ghosh, Managing Founder, illuminis

The Short Answer

CompletionPrism™ scores the prompt, prices every model the company allows from graded study runs, and runs the one with the lowest expected cost for the priority the company chose. Afterwards the receipt shows what was actually spent.

The First Prompt

  1. Score the prompt on 12 signals (length, code, build verbs, scope words and more) and add them up.
  2. Set the size from the score: Simple, Standard or Complex. A conversation with 100,000 or more tokens of context is a long session.
  3. List the models allowed after excluded models, company requirements, budgets and rules. Your own model always stays an option.
  4. Price each model from graded study runs of that size. Take the chance it is still unfinished after each of 3 tries, and the average AI cost of one try. Expected tries are 1 plus those three chances. Expected AI bill is the cost of one try times the expected tries. When AI bill is the priority, human rework cost is left out.
  5. Pick the model. Models that finish less than 80% of the time are dropped, and the lowest expected cost wins.
  6. Show the receipt. The actual bill is the tokens the provider billed times the model's price. Your own model's figure is that actual bill times how much more the study expects your model to cost. The saving is the difference.

Every Later Prompt

Only the size of the conversation is carried forward. The new prompt is priced the same way. Moving a conversation to another model has a cost, because the new model must write the whole context where the current model would only read it at the cheaper cached price. The router moves only when the saving on the turn is at least 1.2 times that cost. Moving is free at the start of a session, so that is the cheap moment to change model.

Three Priorities

Minimize AI bill picks the lowest bill among models above the finish floor. Minimize rework picks the least rework within a spend ceiling. Optimize AI bill versus rework picks the lowest bill plus rework. The priority is shown on every receipt.