A Bot Team Emailed 12 LLM Resellers With a $30,000-a-Month Budget — Only Five Replied
A team that builds Telegram bots decided, after Claude was blocked, to buy LLM access the grown-up way: with a contract, a UPD (Russian universal transfer document) and VAT. Twelve Russian API resellers received the email with a $30,000-a-month budget, and only five replied within four days. Token prices among the finalists differed by less than 2%, but tax terms changed the picture far more.
AI-processed from Habr AI; edited by Hamidun News
When bots bring in money and access to Claude suddenly shuts off, the question of "how to pay for the API" can no longer be solved with a personal card. A team that builds Telegram bots for generating songs, images and business assistants went through this last fall and described it on Habr. The takeaways will be useful to anyone building a product on other companies' models and working from Russia.
How five people ran into accounting
While the team was small, subscriptions and API credits were paid for with foreign cards. In August two things coincided: revenue jumped along with spending on models, and access to Claude and some other models was cut off. On top of that, the team had to change its tax regime: splitting the business to keep tax breaks turned out to be a risky idea.
The decision was to buy tokens like a normal company: under a contract, with a UPD, VAT and no surprises. The same email went to 12 Russian resellers that sell unified access to GPT, Claude, Gemini and other models. It stated a volume of about $30,000 a month (roughly 2.5 million rubles) and asked about token price, VAT, documents, payment terms, API limits and data handling. Within four days five companies replied and seven stayed silent. For the authors, response speed itself became a criterion: if the API goes down at three in the morning, support must also respond quickly.
Why 243 ₽ can be more expensive than 248 ₽
The three finalists — Aporto, Polza.ai and RouterAI — were almost indistinguishable on their price lists. For a million input tokens of the same model they asked 243–248 ₽, and for a million output tokens 1,217–1,241 ₽. The spread is within 2%.
The real difference hides in taxes. An example from the article: an invoice for 1,220,000 ₽ includes 22% VAT, so without tax it is exactly a million. Another invoice for 1,100,000 ₽ without VAT looks 120 thousand cheaper. But if a company is entitled to a full deduction of input VAT, the real cost of the first offer is 1 million, and of the second 1.1 million. The same logic applies to tokens: 248 ₽ with VAT is about 203 ₽ without it, while the competitor's 243 ₽ is tax-free.
A caveat: this is simplified arithmetic. Everything depends on the contract, on your company's right to a deduction, and on whether the reseller sells the service itself or acts as an agent of a foreign supplier. In the latter case, tax agent obligations may arise. It is best to check this with an accountant before signing.
Four criteria and a backup channel
The team chose based on four points:
- a contract with a Russian legal entity and closing documents;
- transparent token pricing: the exchange rate used for conversion into rubles, the markup, and whether there is a top-up fee;
- deferred payment;
- a reserve in case of blocks.
Deferred payment matters because, with spending in the millions of rubles a month, keeping money on a reseller's balance is effectively lending to it.
The authors are sober about blocks: an aggregator itself depends on access channels to OpenAI, Anthropic and Google, so it does not guarantee protection. They kept two providers: a primary and a backup. They did not disclose which ones. Ideally, switching between them comes down to changing a configuration, not rewriting half the application.
What to take away
- Compare the token price after VAT and deduction, not the price list on the website.
- Before topping up the balance, ask the reseller about the exchange rate, markup, fees and its role in the transaction.
- Move the provider into config: base URL and key separate from code, with a unified request format.
- Check how fast support responds in advance, while nothing depends on it yet.
The limitations are obvious: this is one team's experience, the sample of 12 emails is small, the winners' names are hidden, and Russian tax specifics do not carry over to other jurisdictions. But the line of reasoning — documents, real price, backup — suits any business where an LLM has stopped being an experiment and become a line item in the budget.
- Nov 12, 2026Eight Countries, OpenAI and Anthropic Promise an AI "Kill Switch" — but Set No Deadlines or Penalties
- Nov 12, 2026OpenAI and ModRetro Launch a $200 Console for Games Written in Codex
- Nov 12, 2026Hugging Face warns: a wave of hacks could make open-source AI less open
- Nov 12, 2026Anthropic Turns Claude Into a Workspace: Spreadsheets, Whiteboards and Plugins
Key Questions Answered
How five people ran into accounting?
While the team was small, subscriptions and API credits were paid for with foreign cards. In August two things coincided: revenue jumped along with spending on models, and access to Claude and some other models was cut off. On top of that, the team had to change its tax regime: splitting the business to keep tax breaks turned out to be a risky idea. The decision was to buy tokens like a normal company: under a contract, with a UPD, VAT and no surprises.
Why 243 ₽ can be more expensive than 248 ₽?
The three finalists — Aporto, Polza.ai and RouterAI — were almost indistinguishable on their price lists. For a million input tokens of the same model they asked 243–248 ₽, and for a million output tokens 1,217–1,241 ₽. The spread is within 2%. The real difference hides in taxes. An example from the article: an invoice for 1,220,000 ₽ includes 22% VAT, so without tax it is exactly a million. Another invoice for 1,100,000 ₽ without VAT looks 120 thousand cheaper.
What are the key details on Four criteria and a backup channel?
The team chose based on four points: a contract with a Russian legal entity and closing documents; transparent token pricing: the exchange rate used for conversion into rubles, the markup, and whether there is a top-up fee; deferred payment; a reserve in case of blocks. Deferred payment matters because, with spending in the millions of rubles a month, keeping money on a reseller's balance is effectively lending to it.
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.