How a request reaches a model
Software or Bot → the Contractor's server (Russia) → the RouterAI aggregator (Inter LLC, INN 7017351722, Russia) → the model Provider. Some models have a fallback channel — the Timeweb AI service (Timeweb.Cloud LLC, Russia), used only if the main channel does not respond.
Only the request content is sent to the Provider: text, settings, uploaded images or audio. The phone number, Telegram ID, e-mail and key are not transferred — the Provider sees only the Contractor's account. The aggregator may route a request not directly to the developer but through an intermediary cloud (Microsoft Azure, Amazon Bedrock, Google Vertex, DeepInfra and others); in that case that cloud's terms apply to the request. They are shown in the tables.
Training on requests
In the "Training on requests" column:
- no — the Provider's terms for API access do not allow it to use request content to train models;
- yes — the Provider's terms allow it to use requests and results to train and improve its models;
- not excluded — the Provider's terms for such access contain no direct prohibition, or it could not be confirmed.
The user is responsible for the content of a request (Offer, clause 8.4): your own photos and voice are at your discretion; other people's only with their consent. For models marked "yes" or "not excluded", request content may end up in training data. The Contractor seeks to have such use disabled through the aggregator; once it is disabled, the mark will change.
Images
| Model | Developer | Who receives the request (recipient country) | Training on requests | Provider terms |
|---|---|---|---|---|
| GPT Image 2.5 (Sunburst, Flare) | OpenAI | OpenAI, Microsoft Azure or Amazon Bedrock (USA) | no | openai.com/enterprise-privacy |
| Nano Banana Pro, Nano Banana 2, Nano Banana 2 Lite (Gemini Image) | Google, Gemini API (USA); fallback — Timeweb | no¹ | ai.google.dev/gemini-api/terms | |
| FLUX.2 Pro, FLUX.2 Max, FLUX.2 Klein | Black Forest Labs | Black Forest Labs (USA, Germany); fallback — Timeweb | yes | bfl.ai/legal/flux-api-service-terms, bfl.ai/legal/developer-terms-of-service |
| Seedream 5.0 Pro, Seedream 5.0 Lite | ByteDance | BytePlus (Singapore) | not excluded² | docs.byteplus.com/en/docs/ModelArk/Terms_and_Conditions |
| Grok Imagine Image | xAI | xAI (USA) | no | x.ai/legal/terms-of-service-enterprise |
| Krea 2 Large, Krea 2 Medium, Krea 2 Medium Turbo | Krea | Krea (USA) | not excluded | krea.ai/terms, krea.ai/business-terms |
| Recraft V4.1, Recraft V4 Styles (Vector, Pro Vector) | Recraft | Recraft (USA) | no | recraft.ai/legal/developer-terms |
Video
| Model | Developer | Who receives the request (recipient country) | Training on requests | Provider terms |
|---|---|---|---|---|
| Seedance 2.0, 2.0 Fast, 2.0 Mini, 2.5 | ByteDance | BytePlus / Seed (Singapore) | not excluded² | docs.byteplus.com/en/docs/ModelArk/Terms_and_Conditions |
| Kling 3.0 Standard, Kling 3.0 Pro | Kuaishou | Kling AI (Singapore) | yes | kling.ai/docs/user-policy, kling.ai/docs/privacy-policy |
| Veo 3.1 Lite, Veo 3.1 Fast | Google Vertex AI (USA) | no | docs.cloud.google.com/vertex-ai/generative-ai/docs/data-governance |
Sound and voice
| Model | Purpose | Developer | Who receives the request (recipient country) | Training on requests | Provider terms |
|---|---|---|---|---|---|
| MiniMax Speech 2.8 Turbo, HD | text-to-speech | MiniMax | MiniMax (USA) | not excluded | platform.minimax.io/protocol/terms-of-service |
| MAI-Transcribe 2 | speech-to-text | Microsoft | Microsoft Azure (USA) | no | learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacy |
| Nemotron ASR | speech-to-text, fast | NVIDIA | DeepInfra (USA) | no | docs.deepinfra.com/account/data-privacy |
| Lyria 3 Clip, Lyria 3 Pro | music | Google (USA) | no¹ | ai.google.dev/gemini-api/terms | |
| GPT Audio Mini | voice dialogue with the assistant | OpenAI | OpenAI, Microsoft Azure (USA) | no | openai.com/enterprise-privacy |
Language models (Alyosha assistant, chat)
| Model | Developer | Who receives the request (recipient country) | Training on requests | Provider terms |
|---|---|---|---|---|
| GPT-6 Luna, GPT-6 Luna Pro, GPT-6 Sol | OpenAI | OpenAI, Microsoft Azure or Amazon Bedrock (USA); fallback — Timeweb | no | openai.com/enterprise-privacy |
| Claude Sonnet 5, Claude Opus 5 | Anthropic | Anthropic, Amazon Bedrock, Microsoft Azure or Google Vertex (USA); fallback — Timeweb | no | anthropic.com/legal/commercial-terms |
| Gemini 3.1 Flash Lite | Google, Gemini API (USA); fallback — Timeweb | no¹ | ai.google.dev/gemini-api/terms | |
| GLM 5.3 Flash | Zhipu AI (Z.ai) | cloud hosts DeepInfra and others (USA); fallback — Timeweb | not excluded³ | docs.deepinfra.com/account/data-privacy, docs.z.ai/legal-agreement/terms-of-use |
What an answer costs (Offer, clause 9.3)
The cost of a language model answer depends on the length of the request, the accumulated conversation and the answer, so it is known only after the answer and is debited as actually incurred — the amount is shown under each answer. The minimum charge per answer is 0.01 BFT. As a guide — the approximate cost of two typical answers, without discounts for repeated context:
- short — a question with a little conversation (about 1,000 tokens) and an answer of a few paragraphs (about 300 tokens);
- detailed — a long conversation (about 4,000 tokens) and a detailed answer (about 1,500 tokens).
| Model | Minimum per answer | Short answer | Detailed answer |
|---|---|---|---|
| GPT-6 Luna — base model | 0.01 BFT | ≈ 0.02 BFT | ≈ 0.07 BFT |
| GPT-6 Luna Pro — with reasoning | 0.01 BFT | from 0.02 BFT; reasoning can make it tens of times more expensive | from 0.07 BFT |
| GPT-6 Sol | 0.01 BFT | ≈ 0.32 BFT | ≈ 1.29 BFT |
| Claude Sonnet 5 | 0.01 BFT | ≈ 0.32 BFT | ≈ 1.29 BFT |
| Claude Opus 5 | 0.01 BFT | ≈ 0.75 BFT | ≈ 3.12 BFT |
| Gemini 3.1 Flash Lite | 0.01 BFT | ≈ 0.05 BFT | ≈ 0.21 BFT |
| GLM 5.3 Flash | 0.01 BFT | ≈ 0.02 BFT | ≈ 0.05 BFT |
The base model GPT-6 Luna does not debit tokens within the free allowance (Offer, clause 7.6). For image, video and sound generation the software shows the price before starting, and exactly the amount shown is debited.
Notes
¹ Gemini API, Veo and Lyria: Google's terms exclude training on requests under paid access, which the aggregator uses; under free access Google allows such use.
² Under the BytePlus ModelArk terms customer data is not used for training without permission, but this could not be confirmed against the official text of the terms — until confirmed, the mark is "not excluded".
³ Via the aggregator the model is served by third-party hosts. DeepInfra does not use requests for training; the terms of the other hosts are not confirmed.
Timeweb fallback channel: Timeweb's public terms contain no information on whether requests are used for training, so for requests via the fallback channel treat the mark as "not excluded". An earlier version of the same model may answer via the fallback channel.
The list and marks are updated when a new model is connected or a Provider's terms change. Before connecting a Provider from a country not listed in the notification to Roskomnadzor, the Contractor supplements the notification and the Policy (Policy, clause 9.2).