An AI secretary for $77 a month
ИИ-секретарь за $77 в месяц: 88% ответов даёт DeepSeek, а три четверти денег съедают 7% запросов к Opus
B · Useful, limited Solo developer's Habr post with production routing, cost and usage figures from a small Telegram assistant; numbers are concrete but self-reported, small-sample and unaudited.
Summary based on the full text·Original: about 6 minutes to read
In short
A Habr post (12 October 2026) by the developer of "Liza Petrovna", a Telegram and MAX bot that answers in one chat and hands work to internal specialist agents (doctor, lawyer, designer, navigator). The core design choice is routing: a cheap model answers by default and a readable regex plus a length rule decides when to escalate to Opus. The post gives a month of production costs, a topic breakdown of what people ask, and a frank list of failure cases.
Key numbers
- 88%share of the secretary's answers produced by DeepSeek, at $0.002-0.007 per answer
- 7%share of answers sent to Claude Opus 5, which took $39 of the secretary's $51 monthly spend (77%)
- 8568requests from 118 people since 30 May
- 21%share of requests that were small talk or tests rather than tasks
- 4.7%share of real tasks about calendar and reminders, the classic secretary work
As reported by the source.
Russian·Web page·opens habr.com
What it reports
- Last 30 days: $77 in LLM spend across all agents; the secretary alone cost $51, of which $39 went on Opus answers.
- Of 2,500 secretary answers: DeepSeek 88% (2.5-12 s, $0.002-0.007), Opus 5 7% (18-38 s, $0.21-0.24), Haiku 4.5 2.4%, GLM 5.2 2.2% as an experiment. One Opus answer costs about as much as 50 DeepSeek answers.
- Escalation triggers: a regex on words such as "analyse", "compare", "plan", "calculate", "write code", and any message plus context over 2,500 characters. Web search is not a trigger because every model has it.
- Memory: 7,933 stored episodes from 97 people, hybrid vector plus keyword search, mean 33 ms, median 17 ms; falls back to keyword search if the embedding service is down.
- Of 6,796 real tasks: photo or document parsing 19.6%, images and video 12.5%, IT help 9.2%, fact lookup 9.1%; health and legal together almost 12%; calendar and reminders 4.7%.
- Failure cases: answering an earlier photo instead of the new one, wrong reply language driven by a mislabelled metadata field, wrong "today", saving the note "link saved" instead of the link, and misreading domain jargon.
How they got it
Production logs from the developer's own service. Topic shares come from DeepSeek classifying each message into one of 16 topics in batches of 80; the author says no one read the conversations and the shares may be off by a couple of points. Prompt-injection tests are listed with verbatim refusals.
Limits
- Single developer, single small service, one month of cost data; no quality evaluation of cheap versus expensive answers.
- Topic classification was done by an LLM without a checked sample.
- The bot is promoted at the end of the post, and the post received a negative community rating.
Why it matters for companies
The cost profile is typical of assistant workloads: a small share of hard requests drives most of the bill, so routing policy matters more than the default model. A plain, testable rule is easier to audit than a model-based classifier and adds no latency. The usage mix is a warning for anyone scoping an "assistant": people bring photos, health and legal questions, not calendar tasks, so the risky domains need their own guarded paths. The failure list is a practical checklist for agent testing: explicit dates, explicit "new message" markers, and checks that a tool actually stored what it claims.
Key passages in translation
Russian original
Секретарь за 30 дней стоил $51. Из них $39 — это ответы Opus. То есть 7% ответов съедают 77% денег
Over 30 days the secretary cost $51. Of that, $39 was Opus answers. So 7% of answers eat 77% of the money.
Russian original
Да, это регулярка, а не классификатор на модели. Зато её можно прочитать, покрыть тестами и всегда объяснить, почему конкретный запрос ушёл на Opus.
Yes, it is a regex, not a model-based classifier. But you can read it, cover it with tests and always explain why a given request went to Opus.
Russian original
Классической секретарской работы меньше 5%.
Classic secretary work is under 5%.
Summary based on the full text. Read the original for the full argument and data.Read the original on habr.com