Frequently asked questions

Every request is charged for all the tokens the model processed: your message, the whole conversation history, the system prompt, tool descriptions and the answer itself, including the model's reasoning.

Agents — Claude Code, Codex, Cursor, OpenCode and others — re-send the whole context on every step: their own system prompt of thousands of tokens, tool descriptions, open files and history. An agent makes dozens of such requests per task, so spending grows faster than short chat messages suggest.

Re-sent context is charged in full — there is no separate cache discount.

How to spend less: run /compact or start a new session when the context has grown; use a model with a lower coefficient for simple tasks; set a spending limit on the key. Every request and its charge is visible in the dashboard, in the request history.

Before a request runs, the most it can cost is reserved on your balance. The unused part comes back right after the answer.

Your balance is kept in tokens. A request is charged (input tokens + output tokens) × the model's coefficient.

The coefficient is listed in the model catalog and on the pricing page: stronger models have a higher one, light models are below ×1. 1M balance tokens cost $0.05.

Example: a request with 30,000 input and 2,000 output tokens on Claude Sonnet 4.6 (×2.4) is charged 76,800 balance tokens — about $0.0038. The same request on Claude Opus 5 (×6) is charged 192,000 tokens.

Some models bill input and output at different rates — the catalog shows both prices for them. Built-in web search is charged separately, a fixed amount per search.

Yes, they are the vendors' models — OpenAI, Anthropic, Google and others. AnyModel is an independent OpenAI-compatible gateway: you get an AnyModel key, not a vendor account, and one key gives you every model.

How to check: give the same real task to two or three models and compare the results. Each model's availability on real traffic is shown on the status page.

Is it the same model? Yes. Claude is connected through the official Anthropic API and AWS — these are the same models as in claude.ai.

Where does the difference come from? claude.ai and Claude Code add their own system prompt, tool descriptions, extended thinking and conversation context to every request. Together that is tens of thousands of tokens that a plain API request does not have, so the model behaves differently on the same prompts.

Yes, models priced at zero — for example the Free Models Auto Router (am/free): it picks an available free model for every request. It works through the same endpoint and key; only the model value changes. Requests still need at least a small positive balance.

In the dashboard, with the top-up button. Methods: SBP and SberPay, Russian bank cards, crypto (USDT, BTC, ETH) and CryptoBot in Telegram. Which of them are available right now is shown in the payment window.

If the SBP or card payment page does not open, turn off your VPN. The crypto payment page, on the other hand, may not open from a Russian IP.

There is no subscription: tokens are charged only for requests.

Tokens usually arrive within seconds of payment. If the bank confirms a payment late, the service re-checks unconfirmed payments by itself — refresh the dashboard in a few minutes.

If the balance has not changed after 30 minutes, write to support at @ai_anymodel_support with your account email, the amount, and the time and method of payment. It is best to do this within a day of paying.

No. One AnyModel base endpoint — https://anymodel.org/v1 — serves Claude, GPT and the rest of the models.

In the dashboard, under Instructions: provider Codex → Codex → your OS. The installer writes config.toml and auth.json.

Under Instructions: provider Codex → VS Code → your OS. If the models did not appear after install, restart VS Code.

Under Instructions: provider Claude → Claude CLI → your OS. The config is written to ~/.claude/settings.json.

Under Instructions: provider Claude → Claude Desktop → your OS. Use the Gateway URL without /v1 and only the Claude model ids from the generated profile.

Under Instructions, pick the Open IDE provider. OpenCode, opencodex, Zed, Droid, Open Claw, Cline, Kilo Code and Copilot Chat have a one-command installer; Cursor, Windsurf, JetBrains and SDKs get step-by-step manual setup of an OpenAI-compatible connection: base URL and key.

In the model catalog. All models work through the same endpoint — just set the model id you need.

AnyModel is an aggregator: one key and one endpoint give access to every connected model, and all spending comes from one balance.

Through /v1/images/generations, /v1/audio/speech and /v1/audio/transcriptions with the same key.

Restart the app: clients cache the model list. If that does not help, check the base URL and the key.

That the key was copied in full and your balance has tokens. The key is in the dashboard.

Yes. Run the installer again with the new key or edit the config by hand — the old values are overwritten.

No. Use the AnyModel key and endpoint — vendor keys are not needed here.