Is an LLM Matcher Worth It? A Back-of-Envelope

Your pipeline has 300 million candidate pairs after blocking. A LightGBM pair classifier is right on most of them. A teammate proposes sending the pairs to an LLM to decide "same entity?". Use these assumptions (placeholders for this exercise, not real quotes): 300 input tokens and 10 output tokens per pair, USD 1.00 per million input tokens, USD 4.00 per million output tokens, 1 second per call, 32 concurrent calls. Assume 2 percent of candidates (the grey zone, classifier probability between 0.1 and 0.9) are uncertain.

1. Compute the cost and wall-clock time of (a) all candidates and (b) the grey zone only.
2. What accuracy must the LLM have inside the grey zone to be useful?
3. What other number caps the benefit of any better matcher?
4. Recommend an architecture.