The ruble markup on AI tokens

How Russian gateways turned broken payment rails into an access product

On February 24, 2022, Putin's forces officially invaded Ukraine. Not long after, Russia was sanctioned and disconnected from SWIFT, which made payment processors like Visa and Mastercard unusable in Russia. A year later, OpenAI released GPT-4, considered by most to be the start of LLM coding for the technical masses. But Russian developers had a problem: they were unable to access OpenAI's paid service because they were locked out of the global—one may call it Western—financial system. So they turned to another system: a huge resale market built to evade sanctions using proxy accounts and gateways running with foreign credit cards, foreign bank accounts, and LLCs to allow Russian developers and the Russian public to access frontier-level AI. But there was a problem for them: they were a captive market with nowhere else to go, so the market did what the market does best: it raised their prices.

On 23 September 2026, the Bank of Russia's official exchange rate was ₽84.0657 to the dollar. At that rate, Anthropic's published Claude Sonnet 5 price of $2 per million input tokens and $10 per million output tokens converts to roughly ₽168 and ₽841.

But that's what a Russian end user pays to use Claude.

AITUNNEL (a Russian inference reseller that lists Sonnet 5 as one of its available models) lists Sonnet 5 at ₽400 for input and ₽2,000 for output. ProxyAPI lists it at about ₽600 and ₽3,030. The retail price is between 2.4 and 3.6 times the dollar list price converted at the official exchange rate.

The same pattern appears with more expensive models. Anthropic lists Claude Fable 5.1 at $10/$50 per million input/output tokens, equivalent to roughly ₽841/₽4,203. ProxyAPI lists ₽1,580/₽7,900. GPTunneL lists ₽2,000/₽10,000.

This looks like a token markup, but economically, it is an access premium.

The Russian gateways are not simply reselling compute. They are packaging access to a foreign digital service that a customer in Russia generally cannot buy with a Russian card, cannot officially use from a Russian location, and may not be able to pay through an ordinary cross-border bank transfer.

The product is not the token. The product is the path to the token.

Not just due to sanctions

“Russia was cut off from SWIFT” is a useful shorthand but an incomplete description.

SWIFT is a financial messaging network; it does not hold money or settle transactions. The EU initially prohibited SWIFT services to selected Russian banks in 2022 and expanded financial restrictions over subsequent sanctions packages. SWIFT says it disconnected designated Russian entities to comply with EU law. The Council of the EU says the restrictions now cover more than 100 Russian banks and related financial institutions.

That still does not mean every Russian bank is literally disconnected from SWIFT. The payment problem is broader:

• Visa stopped Russian-issued cards from working outside Russia.
• Mastercard suspended its Russian network services.
• Russia is absent from Anthropic's supported regions for commercial API access.
• USD billing, foreign merchant acquiring, provider fraud controls and IP/location checks add separate points of failure.

A developer in Moscow therefore has a different purchasing problem from a developer in Berlin. The question is not merely, “What does one million tokens cost?” It is, “How do I establish a supported account, move rubles into the foreign commercial system, keep the account alive and reach the endpoint?”

Russian gateways collapse those problems into one API key.

The gateway is a financial product disguised as developer infrastructure

The technical integration is almost trivial:

client = OpenAI(
    base_url="https://api.russian-gateway.example/v1",
    api_key="local-gateway-key",
)

The commercial integration is the hard part.

The public offer usually combines:

• payment by Russian card, Faster Payments System transfer or domestic invoice;
• a ruble-denominated balance;
• a reachable Russian or foreign gateway endpoint;
• foreign upstream procurement;
• currency conversion;
• compatible request and response formats;
• local customer support and accounting documents;
• model switching behind one key.

ClaudePoint says Russian clients call its endpoint while provider routing happens on its side. llmgw advertises Russian cards, SBP transfers and invoices without a foreign card or client-side VPN. ProxyAPI publishes prices in rubles with VAT and accepts domestic payment. RuAPI takes the opposite approach and uses USDT or USDC rather than Russian banking rails.

The observable chain looks like this:

Russian customer
  → rubles, SBP, local card, invoice or stablecoin
  → Russian-facing gateway
  → foreign company, cloud account, aggregator or supplier
  → model provider
  → compatible endpoint returned to the customer

Every arrow can add cost. The gateway may pay payment-processing fees, foreign-exchange spread, tax, foreign-company costs, cloud or aggregator markup, fraud losses, support, infrastructure and the risk that an upstream account or route disappears.

The published price spread is therefore not the operator's profit margin. ProxyAPI explicitly says its prices include VAT, while an official dollar list price may exclude local taxes. The gateways also provide services that the model vendor's list price does not include.

Still, a 2.4–3.6× retail multiple is large enough to reveal the value of the missing payment and access rails.

Three Russian token markets

The public pages show three distinct markets operating beside one another.

1. Formal ruble gateways

These services look like ordinary B2B infrastructure. They publish token tables, legal terms, invoices and business support. ProxyAPI, GPTunneL, RouterAI, Polza AI and several others identify a legal operator or named team.

Their central proposition is convenience and continuity: one ruble balance, one integration, many foreign models, no need for each customer to solve cross-border payment and account access independently.

The customer knowingly pays more than the converted provider price because direct purchase is not a realistic alternative given the geopolitical constraints.

2. Access-premium marketplaces

On Plati and similar digital-goods markets, some sellers charge more than one dollar for one nominal dollar of API balance. That is the cleanest expression of the access business. The markup is attached directly to the ability to obtain a working balance or key.

These listings may deliver a direct key, an OpenRouter balance, an activation link, or access to the seller's own gateway. The buyer pays for immediacy and accepts more counterparty risk, with no assurances if that key, balance, or activation link becomes unusable in the future.

3. Gray pool suppliers

Other sellers advertise prices below the original provider's list price—sometimes dramatically below it. These sellers are part of the gray market that are getting lower priced tokens through fixed-price subscriptions, unused enterprise quota, cloud commitments, promotional credits, caching, a seller-defined meter, model substitution, unauthorized accounts or some combination.

The Russian-language Telegram channel TKBK says a nominal dollar of Anthropic limit costs the buyer only $0.10–$0.25 and advertises discounts of up to 97%. It refers publicly to account pools, supplier changes, foreign nodes and downstream reselling.

The payment blockade creates demand for formal gateways. Their high markups create demand for gray suppliers. The two markets reinforce each other.

The markup is visible because the unit is standardized

AI inference is unusually easy to repackage. In fact, even outside of illicit markets, it is constantly repackaged in the West through inference routers like OpenRouter or Together AI. The buyer already expects to receive a bearer key, a base URL, and a token counter. Gateways developed for benign use cases can preserve those interfaces while replacing everything behind them.

The customer sees:

model = "claude-sonnet-5"
input_tokens = 1,000,000
output_tokens = 1,000,000

The customer may not see:

immediate supplier = global aggregator
payment currency = stablecoin
upstream credential = foreign enterprise account
provider relationship = indirect
actual routing = changed after an outage
tax and settlement layers = several intermediaries

This lets a gateway sell a standardized unit similar to what frontier labs and 1st party inference providers sell.

It also creates several incompatible meanings of “token.” Formal gateways commonly publish separate input, output and cache-token prices. Marketplaces may sell house tokens, internal credits, nominal balance dollars or time-limited “unlimited” access. The headline price is comparable only when the model version, tokenization, cache treatment and meter are the same.

Why customers accept the premium

The obvious explanation is that the alternative costs more than the sticker price when the official provider is made inaccessible, much like we saw during Prohibition with bootleggers.

A Russian company that wants to buy directly needs a foreign legal entity or trusted intermediary, a supported-country account, foreign banking, cross-border accounting, stable access infrastructure, and someone to handle blocks and provider changes. A small developer may instead need a foreign card, a working account, and confidence that a prepaid balance will not vanish. That infrastructure costs money to set up and maintain while navigating payment and access restrictions.

A ruble gateway converts those fixed costs into a variable token price.

For light or irregular usage, paying 2–4× per token can be cheaper than maintaining the machinery for direct procurement. For a business, a domestic invoice and closing documents may be worth more than the raw compute. For a developer, changing one base_url is easier than rebuilding a payment and identity stack.

This is the same reason importers can charge large spreads on scarce physical goods. The foreign list price is not the local replacement cost.

Where the economics become opaque

But this formal-gateway explanation has limits.

A markup above list price is consistent with taxes, settlement and service. A price far below list requires a different economic model that is not always 100% clear. Public Russian-language pages document several alternatives:

• EveryAPI invites users to attach unused Claude Pro, ChatGPT Plus, Gemini and corporate-account quota through credentials or OAuth, then shares downstream revenue.
• OmniRouter documents pools of subscription accounts with automatic rotation when one reaches its limit.
• A Plati seller says its capacity comes from pools of Max subscriptions and enterprise cloud accounts.

Those mechanisms turn fixed-price or already-purchased quota into metered retail traffic. They can undercut formal gateways because the marginal cost of another downstream request is not the provider's published API price.

Account pooling does not prove credential theft. An account owner may supply access voluntarily while violating a provider's resale rules. Other accounts may be promotion-farmed, fraud-funded or compromised. The compatible endpoint hides the evidence required to distinguish them.

What sanctions really created

The most interesting result is not that sanctions made software harder to buy. It is the market structure that emerged in response to evade these sanctions to satisfy the demands of their public.

Payment and country restrictions inserted a new commercial layer between Russian developers and model providers. That layer now controls:

• which upstream supplies the model;
• the ruble exchange rate embedded in token pricing;
• the true meaning of the retail token or credit;
• customer identity and local billing;
• failover between suppliers;

The model provider sees a foreign account, cloud customer, or aggregator. The Russian customer sees a domestic invoice and a compatible key. The gateway owns the mapping between them. It also controls who gets access and who does not.

That mapping is not only valuable but it is a great place to hide unauthorized pooling, stolen credentials or model substitution. The same infrastructure that solves a real payment problem can obscure upstream provenance and can allow also foreign actors to run distillation as well as model extraction, as we have seen in the Chinese marketplaces.

The token is the receipt

Russian gateways appear to sell AI tokens. What they actually sell is a completed cross-border transaction.

The token is the receipt for rubles converted into foreign model access, routed through a reachable endpoint and wrapped in local support and paperwork. Its retail price includes the technical service, but the scarce input is settlement.

That is why the same nominal token can cost 2–4× the converted list price at a formal gateway and 80–97% less than list price in a gray pool. Both prices are responses to the same missing direct market.

The price spread is not a curiosity in a token table. It is the visible cost of a missing direct market: payment access, account continuity, routing, and local commercial support have all moved into the gateway. That also means the gateway—not the token counter—becomes the critical point of provenance. When payment rails break, APIs become financial intermediaries.