AI Agent Reply Costs Vary 100x: From 0.94 ₽ to 121.57 ₽ per Message

When a team builds an AI agent, the first question is usually: which model should it run on? The Agents Place catalog analysis suggests a different order. The authors looked at 69 agents and compared what a single reply costs. The spread came out at more than 100x: from roughly 0.94 ₽ per message to 121.57 ₽. Model choice explains only a small part of that gap.

How an agent differs from a chatbot

A chatbot replies with text drawn from what the model "remembers." An agent can do something in the outside world: call an API, run a calculation, assemble a file. It decides on its own when that is needed.

The catalog offers a telling example. Ask "How many people search for coffee delivery in Vladikavkaz?" and an ordinary model will invent a plausible number. The "Demand Analyst" agent goes to Yandex Wordstat, pulls the real search frequency, and builds a seasonality chart right in the reply. Its description says so explicitly: the figures come from Wordstat and are not made up.

The formula in the analysis is simple: an agent is a model, an instruction, tools, and a loop in which the model decides whether to call a tool and what to do with the result. The instruction turns the same model into a "Code Reviewer" or a "System Debugger." Tools work as hands. "Statement of Claim" fills in a PDF, "Instagram Search" looks up profiles and hashtags, and "Video Generation" shoots a clip. Users pay per actual reply, with no subscription. The bill includes tokens at the agent's rate plus paid calls to external services.

How much a single reply costs

Prices were taken from the agent pages on 5 November 2026. The number of tokens per reply in the table is the authors' estimate: the message price divided by the rate.

• Instagram Search (DeepSeek V4.1 Flash, rate 0.0013 ₽ per token): about 0.94 ₽ per message, roughly 720 tokens.
• Statement of Claim (Claude Sonnet 5.5, PDF tool, rate 0.0012 ₽): about 2.80 ₽, roughly 2 300 tokens.
• Code Reviewer and the other IT agents (all 10 run on Claude Sonnet 5, with no external tools): about 2.95–2.97 ₽.
• Demand Analyst (Claude Sonnet 5, Wordstat and charts, rate 0.0025 ₽): about 6.81 ₽, roughly 2 700 tokens. This figure is an average over 10 messages across 30 days; for the other agents, the price shown is a forecast.
• MedScreenwriter (GPT-6 Astra): about 22.69 ₽.
• Video Generation (Claude Sonnet 5, rate 0.0025 ₽): about 121.57 ₽. Tokens barely matter here, because the video price comes from the vendor's response at the time of the call.

The user-created SMM agents cost roughly 20–23 ₽. The catalog's authors suggest the cause is long instructions and elaborate replies. That is their hypothesis; the article offers no direct check.

The table shows three effects. First, the model changes the price by a factor of a few, not by tens. DeepSeek Flash is roughly three times cheaper than Claude, mainly because it gives shorter replies. Second, a tool raises the price by an order of magnitude. The video agent and Demand Analyst run on the same model at the same rate, yet their replies differ in price by 18x. Third, the agent's author also affects the bill, because the instruction determines how verbose the reply will be.

What to do if you build agents

Rule one: calculate the price of a reply before choosing a model. A cheap model suits narrow tasks. Search and text checking on DeepSeek Flash cost less than one ruble, while the catalog assigns contracts and code to Claude.

Rule two: watch reply length. 2 300 tokens versus 720 is already a threefold difference, and the per-token rate does not offset it.

Rule three: treat paid tools as a separate budget line. An agent should ask for confirmation before an expensive call. The video agent does this: it first proposes a prompt, then writes "Confirm, and I'll start the generation."

Rule four: show the price before the request. On every Agents Place card there is a forecast for one message and for 10 messages, so users get no surprises on the bill.

There are caveats. The article was published on Habr by the catalog itself and describes its own agents, so it is closer to a case study than an independent study. The token count is a calculated estimate, and for most agents the price remains a forecast, not a fact. Still, the orders of magnitude and the logic of the calculation are useful to anyone budgeting for their own agent: count the tools and reply length first, and choose the model afterward.