Route the meaning.
Audit the money.
Run real prompts through local MiniLM embeddings, inspect every alternative, then test the workload against your own token-price assumptions.
Prompt batch
ONE PROMPT PER LINE - 2 TO 6Fable 5 / complex coding
GPT 5.6 / reasoning
Grok 4.5 / everyday
Prices are user-supplied scenario inputs, not live quotes. Routing evidence measures semantic fit to the authored anchor set, not model quality, latency, context length, safety, or availability.
Route evidence
MODEL EVIDENCE PENDINGBudget hypothesis awaiting evidence.
The projection will follow from the selected route mix, token volumes, illustrative prices, and monthly workload.
A route is a decision,
not a verdict.
Meaning chooses the lane. MiniLM maps prompts and transparent anchor examples into the same 384-dimensional space. The nearest average anchor score supplies the route.
Margin exposes doubt. A small difference between the best and second-best lane means the router should ask for review, not manufacture certainty.
Usage determines spend. Monthly cost is route mix multiplied by input tokens, output tokens, illustrative prices, and prompt volume. A subscription claim alone cannot establish that total.