> ## Documentation Index > Fetch the complete documentation index at: https://docs.ruapi.ai/llms.txt > Use this file to discover all available pages before exploring further. # Why some sell AI APIs at 5% of the official price — and what it costs you > Where Claude, GPT and Gemini API offers 10–50× below official prices come from, what the buyer risks — outages, hidden system prompts, data leaks, a vanished balance — how to spot such a service, and how to pay less for tokens honestly. **In short:** the provider's official price is the cost floor for any honest reseller. Large customers do get discounts, but not 90–98%. If Claude, GPT or Gemini is sold at 2–10% of the official price, the tokens almost certainly don't come from the official API: they come from resold subscriptions, trial credits, other people's accounts, or a swapped model. It works until it gets shut down — and your balance goes with it. ## How cheap it gets Offers we saw on the market on 27.10.2026, compared with the providers' official prices (\$ per 1M tokens, input / output): | Model | Official price | Seen on the market | Share of official | | - | - | - | - | | Claude Opus 5.5 | 4 / 20 | 0.30 / 0.30 | 7.5% / 1.5% | | Claude Sonnet 5.5 | 2 / 10 | 0.70 / 0.70 | 35% / 7% | | GPT-6 Astra | 10 / 50 | 0.44 / 0.44 | 4.4% / 0.9% | | Gemini 3.8 Flash | 0.75 / 3.75 | 0.04 / 0.20 | 5.3% | > [!NOTE] > > Official prices are from the [Anthropic](https://www.anthropic.com/pricing), [OpenAI](https://openai.com/api/pricing/) and [Google](https://ai.google.dev/gemini-api/docs/pricing) pricing pages. For Gemini 3.8 Flash this is Google's promotional price, valid until 31.12.2026. The same price for input and output is a hint in itself: all three providers charge 4–6× more for output than for input. If the seller doesn't care how many tokens the model writes, they are most likely paying for something other than tokens. ## Where such prices come from | Source | How it works | What the buyer gets | | - | - | - | | A subscription posing as an API | One subscription (Claude Max, ChatGPT Pro) is shared among dozens of users through an unofficial proxy that pretends to be Claude Code or Codex | A breach of the provider's terms and regular bans. To pass as Claude Code, the proxy often adds someone else's system prompt to your request | | Trial credits and free tiers | Accounts with free cloud credits and trial limits are registered in bulk | Access lasts until the accounts are noticed and closed | | Other people's keys and accounts | Stolen API keys, or accounts paid for with someone else's card | Sudden cut-offs, and your requests run through accounts you have nothing to do with | | A swapped model | A cheaper or cut-down model answers under the name of an expensive one | Worse quality that is hard to notice without testing | ## What it costs you * **Outages.** A channel works until the account behind it is banned. Today every request goes through, tomorrow none do, and round it goes. * **A hidden system prompt.** If someone appends their own instructions to your request along the way, the model answers differently: extra tokens, unexpected refusals, a different coding style. * **Your data.** Your prompts and code pass through third parties' accounts. If the service has no legal entity or address, there is no one to hold responsible for a leak. * **Your money.** Your balance lives as long as the service does. If refunds are only given when the whole service is down, everything else is your risk. * **A floating price.** If the price is recalculated every minute, you can't budget a project. ## How to spot it * Flagship models cost a fraction of the official price, and the service doesn't explain where the discount comes from. * The terms and the website name no legal entity, country, address or governing law. * One model has dozens of unnamed "routes" with different prices and reliability, and the price changes every few minutes. * Input and output cost the same, although the provider charges several times more for output. * The only way to reach them is Telegram. ## How to pay less honestly You can save without grey schemes: * **Prompt caching.** Agents and chatbots send the same context over and over. Repeated context read from the cache costs at least 10× less than regular input — see [Prompt caching](https://docs.ruapi.ai/en/prompt-caching.md). * **The right model for the job.** Routine work is fine on inexpensive models — `deepseek-flash`, `claude-haiku-5-5`, `gpt-5.4-mini`, `gemini-3.8-flash`. Keep flagships for hard tasks. * **History and answer length.** Cap `max_tokens` and don't feed the whole project into the agent. How to work it out — see [How to cut your Claude Code and Cline bill](https://docs.ruapi.ai/en/claude-code-cline-costs.md). * **Provider promotions.** Providers cut prices for a while: for example, Google halved Gemini 3.8 Flash prices until 31.12.2026, and RuAPI's prices dropped the same way — see the [changelog](https://docs.ruapi.ai/en/changelog.md). ## How RuAPI does it * Original models — no swapping and no silent downgrades. * Major models have backup channels: if one channel fails, the request automatically goes to another. * Prices are fixed and visible before you pay. They change when the providers' prices change, and we post every change in the [changelog](https://docs.ruapi.ai/en/changelog.md). * Every request is visible in the Console log: model, tokens, charge. * The operator is CYBERBIT PTE. LTD. (Singapore): see the [terms of use](https://docs.ruapi.ai/en/legal/terms.md); an unspent balance can be refunded within 7 days of the top-up — see the [refund policy](https://docs.ruapi.ai/en/legal/refund.md). ## FAQ > **It works, so what's the difference?** > > It works until the accounts your requests go through are banned. For a one-off try that may not matter. If your product or workflow depends on the API, an outage on any given day and a vanished balance cost more than you saved. > **Aren't there enterprise discounts?** > > There are: large customers negotiate volume and prepayment discounts with providers. But such discounts don't turn \$20 per 1M tokens into \$0.30. And a reseller buying at a discount still has to cover its own costs. > **How can I check it's really that model answering?** > > The `model` field in the response proves nothing — the service writes it itself. It's more reliable to test the capabilities the provider advertises: context length, image input, function calling. You can also compare answers to the same hard task with the official chat. > **Why doesn't RuAPI have prices 20× below official?** > > Because we buy tokens legally and can't sell them below cost. You can still save honestly: with prompt caching and the right model, a coding agent often costs several times less than at list price without caching. ## Next > **[Prompt caching](https://docs.ruapi.ai/en/prompt-caching.md)** > > Repeated context at least 10× cheaper. > **[Cut your Claude Code and Cline bill](https://docs.ruapi.ai/en/claude-code-cline-costs.md)** > > What drives the cost and how to lower it. > **[Prices for all models](https://www.ruapi.ai/pricing?lng=en\&utm_source=docs\&utm_campaign=en-cheap-api-risks)** > > Current prices per 1M tokens. > **[How to buy USDT](https://docs.ruapi.ai/en/buy-usdt.md)** > > Crypto Bot, Bybit P2P or an exchanger — three ways. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.