AI Secretary Liza Petrovna Spends $77 a Month: DeepSeek Gives 88% of Answers

The author of "Liza Petrovna", an AI secretary in Telegram and MAX, has posted numbers from the live server on Habr. Since May 30, 118 people have sent the bot 8568 requests. Language models for all agents cost $77 over the last 30 days. DeepSeek gives almost all the answers, 88%. The expensive Opus model gets about 7% of requests but consumes roughly three quarters of the budget.

How the secretary works

Liza Petrovna works as a dispatcher. The user writes in a messenger, and the secretary decides which agent to hand the task to. The author promises that the full article will show the model-selection code and break down the memory built on pgvector, that is, vector search inside Postgres. The published announcement makes the main point clear: the system does not push every request through one expensive model but distributes them by complexity. The announcement gives no details on the routing, so you will have to read the original for those.

Another detail concerns the requests themselves. The calendar, which is what people usually associate with a secretary, accounts for less than 5% of requests. The author reveals in the main text what people use the bot for in the other 95%.

Where the money goes

Let's do the math on the published data. $77 a month comes to about $2.6 a day. Three quarters of the budget, around $58, goes to Opus. That leaves about $19 for everything else, including DeepSeek and its 88% of answers.

If the shares of requests and spending are measured on the same base, the rough estimate is that the average Opus request costs about 40 times more than the average request to the other models. That is an estimate from the two percentages in the announcement, not a measured figure. But the order of magnitude is telling. The savings come not from haggling over token prices but from deciding which requests get access to the expensive model at all.

Key project metrics:

• period since May 30: 8568 requests from 118 people;
• LLM costs for all agents over 30 days: $77;
• DeepSeek's share of answers: 88%;
• share of requests to Opus: about 7%, share of spending: about 75%;
• calendar requests: less than 5%;
• memory: pgvector;
• channels: Telegram and MAX.

What you can take away

The case has limits. It is data from a single author, the audience is 118 people, and it is not a quality comparison of models. It does not show that DeepSeek replaces Opus. It shows only that on a real stream of everyday requests, most answers do not need the expensive model.

For anyone building similar agents, this boils down to a few practical steps:

• Don't track cost as a blanket average; track it by request class. A distribution like "7% of requests, 75% of the money" is easy to miss if you only look at the total bill.
• Make the cheap model the default, and bring in the expensive one only on an explicit complexity signal.
• Look at what users actually ask for. If the calendar takes up less than 5%, other scenarios are worth optimizing.
• You don't have to set up a separate vector database for memory. pgvector runs in the familiar Postgres, and at this scale that is enough.
• Record failures honestly. The author separately covers where the secretary still makes mistakes, and a section like that is more useful than polished demos.

The presence of MAX alongside Telegram is also a signal: Russian bots increasingly launch on several messengers at once, and model routing becomes a shared layer across all channels.