What 6 cents of meeting analysis actually does

Six cents

$0.06. That's roughly what an average client meeting costs in model fees in the analysis pipeline we built. For that, an agency gets a summary, a numbered list of decisions, draft action items, every request the client made with their exact words, and a scope check on each request against the signed proposal.

The number is small because of a design choice, not because models are free. Here's how the work is split, and why.

Two jobs, two models

Reading a transcript and judging scope are different jobs. Reading is long and shallow: thousands of words of conversation, of which a handful matter. Judging is short and deep: a few requests, checked carefully against a proposal and against what was decided in earlier meetings.

So the pipeline uses two models. A fast, inexpensive reader goes through the whole transcript once and extracts the summary, the decisions, the action items and every client request with its quote. It makes no judgments. Then a stronger model sees only those requests, the summary, the signed proposal and earlier meetings. It never sees the transcript.

The split protects quality as well as cost. A model asked to read an hour of conversation and judge scope in the same pass has to hold everything at once. A judge that sees a clean list of requests next to the proposal's line items has one narrow question to answer for each: in scope or out, why, and roughly how many hours.

The price list explains the split

The economics are on the providers' own price pages. Gemini 3.6 Flash, the reader, costs $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, rising to $1.50 and $7.50 from January 1, 2027. Claude Sonnet 5.5, the judge, costs $2 per million input tokens and $10 per million output tokens.

The judge's higher price matters much less when its input is a short list of requests instead of an hour of conversation. The long reading happens on the model priced for long reading.

The cheapest call is the one you skip

If the reader finds no client requests, the judge isn't called at all. Plenty of status meetings end there: summary, decisions, action items, done. No scope question means no judge cost.

If an agency has no AI key, a rule-based reading runs instead of the two models. It's less nuanced, but the meeting page still works.

Input is capped too. The reader takes at most 400,000 characters of transcript, so a marathon workshop can't quietly turn into an expensive call.

What the money buys

The output is a draft, not a decision. Action items wait as drafts until someone approves them, and unapproved ones never reach a board or a client. A scope alert shows the client's words, the speaker, the reason the request falls outside the proposal and the estimated hours.

The amount on that alert is hours times the project's rate, computed in code and rounded, never generated by a model. Language models are good at reading what a client meant. They shouldn't be trusted with arithmetic that ends up on an invoice.

Every model call is logged with its model, its tokens and its cost in dollars. So six cents isn't a marketing estimate; it's a number an agency can check against its own usage.

Where SmoothOS fits

This is the meeting pipeline inside SmoothOS. Today it runs on speaker-labeled transcripts delivered to its intake endpoint, and a recording bot that joins your calls is coming soon. Each analysis is logged with its model and cost, the last 30 days of AI spending are visible in Settings, and nothing reaches a client until a person approves it.