Back-of-the-envelope math: splitting tasks across models cuts a contact center's bill from 207,000 ₽ to 23,000 ₽
If both email sorting and reply drafting go to one expensive model, the AI bill climbs without any gain. In a calculation for a contact center handling 50,000 requests a month, running everything on Claude Opus 5 costs about 207,000 ₽, on Sonnet 5 it's 83,000 ₽, and with the steps split across models it drops to about 23,000 ₽. A cheap model extracts fields, a mid-tier model writes the draft, and a strong model checks 25% of the flow. The expensive model is reserved for complex cases.
AI-processed from Habr AI; edited by Hamidun News
Picture a contact center handling 50,000 requests a month. The AI pilot ran on the most powerful model, and people liked the results. Then the same model was applied to everything: identifying the topic of an email, pulling out an invoice number, drafting a reply, checking quality.
A month later the bill had grown several times over, even though a cheap, fast option could have handled most of the work. This is the subject of a Habr article on building a multi-model pipeline that turns 207,000 ₽ a month into roughly 23,000 ₽. A caveat: the text was written by the team behind the Russian gateway Heli and partly promotes it, and the figures come from a worked example, not from client data.
Where the one-vendor, one-model setup breaks down
The first problem is money. An expensive model performs tasks that a cheap one could handle. The bill grows not because companies use AI more, but because simple classification is priced like complex reasoning.
The second is procurement. If access is contracted directly with a single model developer, swapping any step means a new contract, a new offer, or a new limit. If employees rely on personal subscriptions, the finance director has no act, no department-level limits, and no log. And when the vendor goes down, the entire pipeline stops.
The third is checking with the same model that wrote the draft. An error in the model's own style easily slips through. It is better to give the second pair of eyes to a different model and leave the final decision to a human.
Which model does what
The example has five roles, with prices in rubles per 1 million tokens (input / output):
- Extraction and classification: DeepSeek V4 Flash, 11.5 / 22.5 ₽. It needs stable JSON, speed, and a low price.
- Reply draft: Gemini 3.7 Flash, 52 / 258.5 ₽. It assembles coherent text from a policy document or knowledge base.
- Review: Claude Sonnet 5, 276 / 1,378.5 ₽. It does not rewrite the text. It looks for contradictions, invented facts, off-tone wording, and missed risks.
- Fallback: GPT-5.4 mini, 103.5 / 620.5 ₽. If the primary model returns a 5xx error or hangs, the request is retried with a different vendor.
- Baseline for comparison: Claude Opus 5, 689.5 / 3,445.5 ₽. This is the "everything on one flagship" scenario.
For claims, contracts, personal data, and money, a human remains at the output stage. The model itself sends nothing to the client.
In code, this is a standard OpenAI-compatible client with the gateway address and a single key. Each step maps to a model identifier, and if one fails, the request goes to the fallback identifier. The company keeps the routing logic in-house, in its service layer, message bus, or n8n. The gateway provides unified access, balance management, and closing documents.
What it costs
The authors budget three operations per request: classification (800 input tokens and 200 output), drafting (2,500 and 350), and spot-checking 25% of the flow (2,000 and 250). Token counts are estimates, not measurements from production logs. Prices come from the gateway's public catalog as of November 13, 2026.
- Everything on Claude Opus 5: about 207,000 ₽ a month.
- Everything on Claude Sonnet 5: about 83,000 ₽.
- Routed pipeline (DeepSeek V4 Flash for classification, Gemini 3.7 Flash for drafts, Sonnet 5 for 25% of reviews): about 23,000 ₽.
The difference is three to nine times. Review by a second model does not double the bill if it is applied not to the whole flow but to complaints, large deals, high legal risk, and disputed cases.
What the access layer needs beyond price
For the pipeline to run in production, you need a ruble invoice and an act based on actual consumption, separate keys and limits per department, a model selection policy, and an expense log. It is best not to store request texts in the log: for accounting and investigations, metadata is enough, namely the model, token counts, and cost.
Heli offers this: a shared balance, projects as cost centers, limits per day, week, or month, email alerts at 50, 80, 90, and 100% of spend, roles for finance staff, CSV or JSON export, and a monthly act. The balance is prepaid: if funds run out, the API returns error code 402 and the model request is not sent. The authors name the limitations themselves: invoices are issued without VAT, and UPD and EDI are not yet available. This is worth clarifying with accounting before the first payment.
Legally, the contract is with the gateway, not with OpenAI, Anthropic, and Google at the same time. The point is that the setup is not tied to a single model or a single console.
A 30-day plan
Week 1: pick one flow with a documented policy and measurable volume, then manually label several hundred examples with personal data masked. Week 2: build the three steps and compare them with a run where everything goes through the flagship, on quality and cost. Week 3: enable the fallback model, limits, and logging, and test what happens when the primary model fails. Week 4: show the finance director two numbers, the cost of the routed pipeline and the cost of the same volume on the expensive model, plus the share of drafts that people end up editing.
The practical takeaway: prices fluctuate, so re-check them before setting a budget. The principle doesn't change: use the expensive model for exceptions, let the cheap one handle the routine input, and a single key only removes an unnecessary procurement cycle.
- Nov 21, 2026Shield Hands GPT-6 Astra Work in Legacy Portals Without APIs: Agents 35–40% Faster
- Nov 21, 2026Val Town built a conference site in 3 minutes, while v0 took 15: testing six AI services
- Nov 21, 2026Kimi K3 and DeepSeek V4 Weights Show Traces of Banned Nvidia Blackwell Chips
- Nov 21, 2026AcpGram Lets You Control Claude Code, Codex, and Gemini CLI Directly from Telegram
Key Questions Answered
Where the one-vendor, one-model setup breaks down?
The first problem is money. An expensive model performs tasks that a cheap one could handle. The bill grows not because companies use AI more, but because simple classification is priced like complex reasoning. The second is procurement. If access is contracted directly with a single model developer, swapping any step means a new contract, a new offer, or a new limit. If employees rely on personal subscriptions, the finance director has no act, no department-level limits, and no log.
Which model does what?
The example has five roles, with prices in rubles per 1 million tokens (input / output): Extraction and classification: DeepSeek V4 Flash, 11.5 / 22.5 ₽. It needs stable JSON, speed, and a low price. Reply draft: Gemini 3.7 Flash, 52 / 258.5 ₽. It assembles coherent text from a policy document or knowledge base. Review: Claude Sonnet 5, 276 / 1,378.5 ₽. It does not rewrite the text. It looks for contradictions, invented facts, off-tone wording, and missed risks. Fallback: GPT-5.4 mini, 103.5 / 620.5 ₽.
What it costs?
The authors budget three operations per request: classification (800 input tokens and 200 output), drafting (2,500 and 350), and spot-checking 25% of the flow (2,000 and 250). Token counts are estimates, not measurements from production logs. Prices come from the gateway's public catalog as of November 13, 2026. Everything on Claude Opus 5: about 207,000 ₽ a month. Everything on Claude Sonnet 5: about 83,000 ₽.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.