Hello, fellow arbitragers! My name is Timur, and I am responsible for the marketplace at Cloud.ru. We provide a platform for AI services that integrate into clients’ cloud infrastructure. My task is to ensure pricing transparency. Currently, many evaluate AI by token cost, but this is fundamentally wrong. The client pays not for tokens, but for a specific result. In this article, we will analyze how to correctly calculate the unit economics of an <a href="/ru/blog/massovyy-posting-video-4-strategii-povtoreniya-zadach-ii-agentami/" title="Mass video posting: 4 strategies for repeating tasks by AI agents">AI agent, based on Cloud.ru data, and determine the real cost of a useful response to understand what CPM is profitable.
“A vendor on the marketplace must set a price for their agent, the client must understand it, and the platform must make it possible to compare prices on the storefront. As long as all three parties measure cost in tokens, none of this works.”
Previous materials were dedicated to OpenClaw: how to deploy an agent without data leakage and why a new version is a separate service. Today we talk about money.
Evaluating the Cost of AI Services: Focus on Unit Economics
Unit of Measurement: Useful Response, Not Tokens
A token is a bad metric. The client buys not tokens, but a solution to their problem. A “closed dialogue” is also deceptive. Intercom, for example, distinguishes two types of solutions: confirmed (the client said “thank you”) and assumed (the client simply left).
Both are paid for, but there’s a nuance: if the client returns with the same question, the solution is annulled. An example from Lorikeet (195 cases, 38 vendors): out of 4200 dialogues closed by AI, 900 were closed with a link to a knowledge base article. Without them, the share of resolved issues drops from 70% to 55%.
The difference between “closed” and “resolved” is huge. Therefore, our unit of measurement is a useful response. This is a dialogue closed by AI without human intervention, where the client did not return with the same question within 72 hours and did not leave a negative rating. Everything that did not pass these filters goes into the numerator of costs, but not into the denominator of useful responses. Unsuccessful attempts are paid for by successful ones. This is a key point.
Model Case: Online Store Support
For calculations, let’s take a typical scenario: online store support. This is a mass case with a large volume of public data. The store receives 40,000 inquiries per month. 90% of them (36,000 dialogues) are routed to the AI agent.
Dialogues closed by the agent without an operator, let’s call them auto-closures – this is the metric that vendors show. Agent structure: RAG (knowledge base search + response generation) plus the “order status” tool. One dialogue accounts for three LLM calls:
- Understand the request and choose a route.
- Respond based on the found fragments.
- Refer to the tool or compile a summary for the operator.
On average, this is 12,000 input tokens (system prompt, fragments, history) and 800 output tokens. There are two knowledge base queries per dialogue. The filter model checks the client’s reply and the agent’s response (1000 tokens for each of the three calls). Another 15% of tokens are allocated for re-generations.
Infrastructure and Costs
We use Cloud.ru services:
- Knowledge Base: Evolution Managed RAG.
- Agent: Evolution AI Agents.
- Logs: Evolution Managed PostgreSQL.
Cost of a support operator in Moscow: 125,000 RUB/month (85,300 RUB salary according to hh.ru for 2025 + contributions and overheads). With 40 tickets per day, one ticket costs 149 RUB. Operator minute – 15.5 RUB (125,000 RUB / 80% of 10,080 minutes). For chat in retail (60 requests/day), the ticket cost is 99 RUB.
Assumptions for Calculations
Key assumptions:
- Share of auto-closures: median for public cases – 70%. Ozon: 27% (2020) -> 62% (2022). We take the middle – 50%.
- Dialogues that did not pass the filter: client returned within 72 hours or gave a negative rating. Assumption: 6% (in the Lorikeet example, 21% of closed dialogues were links).
- Cost of such a dialogue: 150 RUB compensation + 3 minutes of operator time (200 RUB).
- Operator time for escalation: 1.5 minutes beyond a regular ticket.
- Manual check: 3% of all dialogues for 3 minutes each.
- Knowledge base, prompts: 0.5 FTE specialist (optimistic).
- Depreciation: 50,000 RUB/month (for an integrator with CRM – up to 1.5 million RUB).
Prices for Models and Infrastructure (Cloud.ru, October 2026)
Evolution Foundation Models tariffs (with VAT 22%, per 1 million tokens):
- HiveTraceGuard-Pro (filter): input 25 RUB, output 125 RUB.
- Qwen3-30B-A3B: input 10 RUB, output 50 RUB.
- GigaChat-3-Pro: input 25 RUB, output 125 RUB.
- Claude Haiku 4.5: input 100 RUB, output 500 RUB.
- Claude Sonnet 4.6: input 360 RUB, output 2250 RUB.
Difference between cheap and expensive model: 36x for input, 45x for output. Infrastructure (Evolution, September 2026, 730 hours/month):
- 4 vCPU + 8 GB RAM: 1.5616 RUB/vCPU-hour, 0.3904 RUB/GB-hour.
- Evolution Managed RAG (2 GB, 72,000 requests): 1.3664 RUB/GB-hour, 0.0976 RUB/1000 requests.
- Evolution Managed PostgreSQL (2 vCPU, 4 GB RAM, 100 GB SSD): 1.49145 RUB/vCPU-hour, 0.4026 RUB/GB-hour, 0.016836 RUB/GB-hour.
Total infrastructure: 13,420 RUB/month. This is the cheapest line in the calculation.
Calculation of the Cost of a Useful Response
Cost of one dialogue
First, we calculate the cost of one dialogue using different models:
- Qwen3-30B-A3B: 0.43 RUB
- GigaChat-3-Pro: 1.32 RUB
- Claude Haiku 4.5: 3.76 RUB
- Claude Sonnet 4.6: 10.99 RUB
A 25-fold difference! If we stopped here, model selection would be critical.
Cost of a useful response
Out of 36,000 dialogues, with 50% auto-closures, 18,000 were closed. Of these, 6% were errors. This leaves 16,920 useful responses. Each useful response incurs not only token costs but also:
- Tokens from failed attempts.
- Operator minutes on escalations.
- Cost of errors.
- Verification.
- Knowledge base.
- Implementation.
The 25-fold difference in tokens translates to a 1.4-fold difference in useful response cost: from 51 RUB to 73 RUB. The expensive model (Claude Sonnet 4.6), which is 25 times more expensive in terms of tokens, turns out to be only 40% more expensive in terms of result cost. Why? Because all other costs are the same for all models.
Price structure: GigaChat-3-Pro (53 RUB)
For GigaChat-3-Pro (53 RUB per useful response):
- Escalations and erroneous responses: 71% of the cost. These costs are usually not accounted for as “AI expenses” but are embedded in support salaries or compensation.
- Tokens: 2.8 RUB (5%). This is the only line item everyone compares.
- Cloud infrastructure: 0.80 RUB.
For Claude Sonnet 4.6, tokens already account for 32% of the price. Here, it makes sense to save on tokens. How?
- Context reduction: reduce 12,000 input tokens to 6,000. The useful response on Sonnet becomes cheaper from 73 to 65 RUB.
- Caching: reading from cache is 10 times cheaper. If half of the context is a system prompt, the price drops to 66 RUB.
On Qwen, these techniques save pennies.
Key levers for savings
1. Auto-closure rate
Transitioning from a pilot (35-50% auto-closures) to a production environment (65%) saves 15-19 RUB. Ten percentage points of auto-closures yield as much as switching to a cheaper model (on Sonnet, an increase from 50% to 60% saves 14 RUB, replacing Sonnet with Haiku saves 15 RUB). But the former works on any model, the latter – only once.
2. Error rate
Few measure the percentage of closed dialogues that failed the filter, but it is critical.
| Percentage of failed filters | Cost of useful response | Expenses for them (RUB/month) |
|---|---|---|
| 6% | X | Y |
| 3% | X – 8 RUB | Y – Z |
Reducing errors from 6% to 3% saves 8 RUB on each useful response. Switching from GigaChat-3-Pro to Qwen3-30B-A3B saves only 2 RUB. 15% errors can cost more than half of an operator’s salary if every dialogue where the client went silent is considered resolved.
3. Escalation quality
I allocated 1.5 minutes of operator time for escalation (reading summary, continuing). This works with a “warm” handover: the agent provides the operator with a concise history, order number, and list of actions. With a “cold” handover (the operator reads the entire dialogue where the bot answered incorrectly three times), 1.5 minutes turn into 3 or 6.
At 3 minutes, a useful response on GigaChat-3-Pro increases in price from 53 to 77 rubles, at 6 minutes – up to 127 rubles. This is already more expensive than a regional operator. If the agent identifies the client and pre-fills the ticket, the operator saves a minute or two, and this line goes into the red.
“Transmission quality turns out to be a separate budget item with the widest spread in the entire model, and it should be measured: the average processing time for escalated tickets should be calculated separately from regular ones.”
Fighting for accuracy = burning tokens? If, to reduce errors from 6% to 3%, tokens per dialogue increase by 1.6 times, on GigaChat-3-Pro, a useful response still becomes cheaper from 53 to 46 rubles. On Sonnet, extra tokens are more expensive than the accuracy gained; it’s better to improve the knowledge base there.
Conclusions: what to consider a profitable CPM
Practical conclusion: in the working support circuit, the most expensive thing is not tokens, but the denominator. Everything that increases the proportion of useful responses (knowledge base quality, correct routing, test sets) pays off faster than any bargaining for the price of a million tokens.
Expensive models (e.g., Sonnet) pay off if they bring 10 points of auto-closures and 2 points of accuracy compared to cheaper ones. If the difference on your test set is less, you are overpaying. If it’s more, you are unnecessarily saving on tokens.
Comparison with a human: two caveats
1. Ticket price: at 149 rubles per ticket, even a pilot with 35% auto-closures on Sonnet (110 rubles per useful response) looks profitable. Parity with an operator is at 20-27% auto-closures. But if an operator closes 60 tickets a day (99 rubles/ticket), the same pilot on Sonnet (110 rubles) is already more expensive. A regional/outsourced operator (64 rubles/ticket) changes the picture: GigaChat-3-Pro (42 rubles) is cheaper, Sonnet (62 rubles) – parity. The question of “will it pay off” depends on your ticket cost and the proportion of auto-closures.
2. Volume: fixed costs (knowledge base, implementation, infrastructure) – 163,000 rubles/month. For 36,000 dialogues. For 6,000 dialogues, a useful response on GigaChat-3-Pro is 101 rubles, for 3,000 – 159 rubles (more expensive than an operator). Below 3,500 dialogues per month, the agent does not pay off. If fixed costs are higher (a full-time position for the knowledge base – 200,000 rubles/month, implementation – 1.5 million rubles), the break-even point shifts to 7,000 dialogues per month. The share of tokens in this case drops to 4%: the more realistically you calculate other costs, the less significant the price per million tokens appears.
What is not accounted for in the model (in both directions)
- Against the agent: ticket system or chat platform license.
- In favor of the agent: staff turnover (up to 50% per year), recruitment and training costs, slow work of newcomers, scaling for sales, payment for night/weekend shifts. These things are difficult to calculate without specific data, but businesses are often willing to pay for them, even if daily unit economics show parity.
Frequently Asked Questions
What is a “helpful answer” in the context of an AI agent?
A helpful answer is a dialogue closed by AI without human intervention, where the customer did not return with the same question within 72 hours and did not leave a negative rating. This is a key metric for evaluating the real effectiveness and cost of AI services.