We, a team of Telegram bot developers, faced a paradox: despite being ready to spend $30,000 monthly on LLM API, most Russian providers ignored us. This case is about how unit economics and fiscal discipline clashed with the market’s unwillingness to work with real businesses.
From «gray» schemes to white business: why did the problem arise?
Initially, we operated as a typical startup: buying subscriptions and API credits with foreign cards. This was convenient as long as volumes were small. But with growing revenues and API expenses (we won’t discuss profitability, it exists), the accountant naturally began to demand transparency: contracts, closing documents, VAT.
August brought two problems: turnover growth to several million rubles per month and blocking access to Claude and other models. Dozens of bots, growing revenue – and a sudden cutoff from a key service. This was a signal to switch to official cooperation.
Finding a provider: is $30,000 a lot or a little?
We decided it was time to buy access to neural networks like adults: with contracts, VAT, no surprises. The Russian market is full of offers for a unified API to GPT, Claude, Gemini. Beautiful websites, ruble payments.
We drafted one email, indicating a potential monthly volume of $30,000 (about 2.5 million rubles) and sent it to twelve providers. The request was simple: token cost, VAT, closing documents, payment terms, API limitations, data processing.
Market reaction: 5 out of 12 responses – is that normal?
After four days, only five companies responded. The other seven ignored us. A budget of 30 million rubles a year is not cosmic money for the global AI industry, but for a Russian API intermediary, it’s a serious reason for dialogue.
Communication speed is a critical factor, especially when the operation of dozens of products 24/7 depends on the API.
Key requirements for API providers
From the five companies that responded, we selected finalists and detailed our requirements. For us, not only technical aspects but also financial transparency were important.
1. Contract, documents, and VAT: accounting dictates the terms
The accountant was categorical: no «tokens for neural networks.» We needed a contract with a Russian legal entity, invoices, UTD (universal transfer document), and a clear tax scheme. VAT was especially important. From 2026, the VAT rate in the Russian Federation will be 22%.
An offer for 1,220,000 rubles (with VAT) might be more profitable than for 1,100,000 rubles (without VAT) if there is a right to full deduction. One should compare not the nominal price, but the economic cost after tax accounting.
2. Clear Token Cost: Pricing Transparency
We understand that intermediaries need to earn money. But we wanted to see exactly how much we were paying above the official price of the model developer. Exchange rates, markups, VAT, commissions – all of this should be transparent.
Goal: to understand why a million tokens cost differently from different providers.
3. Post-payment: We don’t want to credit the supplier
The “first top up your balance” scheme is suitable for small volumes. With expenses of several million rubles per month, constantly keeping money on the intermediary’s balance is unprofitable. Post-payment is a significant advantage, allowing you not to freeze working capital.

4. Reserve in case of blockages: risk diversification
A Russian aggregator is not a panacea for blockages; it itself depends on access channels. Therefore, we decided to use two suppliers: main and backup. Switching between them should be quick and painless. This minimizes the risks of downtime and loss of ROI.
Comparison of finalists: price vs. economic benefit
Aporto.tech, Polza.ai, and RouterAI made it to the final. To compare them by price, we took one model and the cost of a million input and output tokens:
- 1 million input tokens: Polza.ai – 244 ₽, RouterAI – 243 ₽, Aporto – 248 ₽.
- 1 million output tokens: Polza.ai – 1 218 ₽, RouterAI – 1 217 ₽, Aporto – 1 241 ₽.
The difference is within 2%. But the accountant showed that an API for 243 rubles without VAT can be more expensive than an API for 248 rubles with 22% VAT (if the price without VAT is 203.28 rubles and there is a right to deduction). It is important to consider who the intermediary is (seller or agent) – this affects tax obligations.
“Don’t look at the price from the website, but at the cost after taking into account the tax conditions of a specific transaction” — our accountant.
Selection results and conclusions
We chose two suppliers (we are not disclosing names, we were not paid for advertising). For the main supplier, the critical factors were:
- Final cost including VAT and contract terms.
- Post-payment option.
- Transparent pricing scheme.
- Availability of necessary models.
For the backup, the criteria were less strict, but it had to ensure quick switching. The main conclusion: you cannot build all infrastructure on a single access channel. And yes, the accountant was right. Always.
Frequently asked questions
What CPM should be considered profitable when purchasing AI API?
A profitable CPM depends on your unit economics and the right to VAT deduction. It is important to compare not nominal prices, but the economic cost after taking into account all taxes and contract terms. Price vs. result is the key indicator, not just CPM.
Is an AI API worth its cost if there are free alternatives?
For businesses with growing turnovers and a need for stability, an official API with a contract, VAT, and support is an investment in reliability. Free or “grey” solutions offer no guarantees and can lead to the halt of critical processes, which will ultimately cost more.
What’s cheaper: banner advertising or an API for content generation?
These are incomparable things. Banner advertising is an acquisition channel, an API is a production tool. You should compare the cost of content generation via API with manual production or outsourcing. Cheap hypothesis testing through AI can be more effective than expensive campaigns with unclear ROI.
How to choose an AI API provider for a large business?
Key criteria: transparent pricing (including VAT, exchange rates), availability of a contract and closing documents, possibility of post-payment, operational stability, and availability of backup channels. Ignoring these factors will lead to financial and operational risks.
Can you save on AI API by using multiple providers?
Using two providers (primary and backup) is not so much about saving as it is about risk diversification. This ensures business continuity and protection against blockages, which in the long run helps avoid significantly greater losses.
