Conversational Analytics in Looker: Token Billing From 1 October 2026

Conversational Analytics in Looker is free until 30 September 2026. From 1 October the meter starts: three dollars per million input tokens, twenty dollars per million output tokens, charged as overage against a monthly allowance that each user carries.

The interesting number is not either price. It is the ratio between the two allowances, and what that ratio implies about which questions are worth asking in a chat box rather than building as a report.

The allowance is per person, not per pipeline

This is the structural difference from every other token bill. An API key has usage; a report user has a role. Here the monthly allowance is attached to the role: one million input and twenty thousand output tokens for a Viewer, two million and forty thousand for Standard, four million and eighty thousand for Developer.

Which means the cost of the feature scales with how many people use it, not with how much data it touches. Twenty analysts asking occasional questions can sit comfortably inside the free allowance. One analyst automating a daily narrative summary for twenty dashboards will not, and no amount of query optimisation changes that, because the expense is in the writing rather than the reading.

Why output is the tight half

allowance ratio     1,000,000 in  :  20,000 out      50 : 1
price ratio         3 $/M in      :  20 $/M out    1 : 6.7

  combined: an output token is scarce fifty times over
  and expensive nearly seven times over

cheap question      scans a wide table, returns a number
                    heavy on input, trivial on output

expensive question  "summarise what changed this month
                    and explain why" - light on input,
                    every word of the answer on output

The instinct from other token bills is to worry about how much data goes in. Here that instinct points the wrong way. Input is the generous side both in allowance and in price; the constraint sits on the answer.

A rough sense of scale helps. Twenty thousand output tokens is on the order of sixty to eighty substantial answers in a month for a Viewer, assuming a few hundred tokens of prose each. That is fine for someone who asks a question when something looks odd. It is not fine as the interface to a daily routine.

What this changes about how to use it

The division that falls out of the pricing is close to the division that was sensible anyway. Ad hoc questions, asked because a chart looked wrong, are exactly what a conversational interface is for and they fit inside the allowance without anyone thinking about it.

Recurring output belongs in a report. A narrative that gets generated every Monday morning for the same six dashboards is a template with numbers in it, and a template does not need a language model to fill in once its shape is fixed. Paying twenty dollars a million to regenerate the same sentence structure fifty-two times a year is the failure mode worth naming in advance.

The middle case is genuinely a judgement call: an occasional written summary for a stakeholder who will not open Looker themselves. That has real value, it costs real output tokens, and whether it is worth it depends on the stakeholder rather than on the price sheet.

What to do before October

Two things, neither of them large. First, find out what current usage actually is while it is still free, because after 1 October the same behaviour arrives as a line on a bill and nobody will remember what the baseline was.

Second, look for anything already running on a schedule. Scheduled generation is where a free feature quietly turns into a recurring charge, and the month to find it is the one before the meter starts rather than the one after.