How to buy AI tokens and an API key: 3 methods in 2026

APIs do not have an "all-inclusive" plan: you pay for volume — how many tokens the model has read and how many it has written. Output is almost always more expensive than input, usually three to six times. Therefore, the same task costs differently: a long answer is more expensive than a short one, and a coding agent that sends the project code to the model again at each step spends mainly input tokens.

Balance, not a subscription. ChatGPT Plus or Claude Pro subscriptions pay for chat in the browser and app, while an API is a separate product with its own account. An API key works as long as there is money in the balance: for most providers, this is a prepaid model — top up first, then make requests.

Key. It is created in the provider's personal account, and it is shown in full only once — save it immediately. There can be multiple keys: a separate one for each project or employee to track usage for each and revoke one without affecting others. A key is access to your money: do not post it in code on GitHub and do not send it in chats.

Limits. Providers limit the number of requests and tokens per minute. OpenAI's limits grow with the total purchase amount: the first paid tier opens after $5, the next after $100. For agents and services with high traffic, limits are more important than price: a cheap key that responds "too many requests" during peak hours will not speed up work.

API address (base URL). Most providers, including aggregators and gateways, use the OpenAI format. Therefore, you usually change the provider with two settings — base URL and key, without rewriting the program.

Prices are per 1M tokens, input / output, as of October 2026. Official providers list their flagship models, while JoinGonka offers a unified price for the network's models.

Key takeaways. The official provider is the most direct path to their flagship model, but you must pay in advance, by card, in USD. An aggregator provides hundreds of models under a single key and charges a fee for each top-up. Resellers accept familiar local payment methods, but rarely display the price per 1M tokens, making it difficult to compare their offers with official ones.

Price difference. One million input tokens on JoinGonka costs $0.0069 — about 720 times cheaper than GPT-5.5. One million output tokens — $0.021, about 1,200 times cheaper than Claude Opus 4.8. Note that the models are different: JoinGonka offers open models, while OpenAI and Anthropic offer their own proprietary flagships. Open models are usually sufficient for chatbots, automation, document search, and coding agents; if a task requires specifically GPT or Claude, it is more cost-effective to buy access directly from the developer.

How to choose. If you need a specific flagship model (GPT or Claude), buy directly from its developer. If you need many different models under one key and don't mind the top-up fee, an aggregator is suitable. If the price per token matters and open models suffice for your task, a gateway to an open network is more profitable. A reseller is only justified if the seller passes the verification from the next section.

Sellers of "AI tokens" and ready-made API keys are reselling access to third-party models. There are two common schemes. In the first, the seller provides a key to their own server, which then forwards requests to OpenAI, Anthropic, or other providers. In the second, multiple buyers gain access to a single official account or subscription. Low prices are often explained by the fact that a large plan is split among many buyers, which is where the risks come from.

• Shared key. If many people use the same key, you share the rate limits with them: during peak load, requests hit "too many requests" errors, and someone else's usage is deducted from your shared balance.
• Banning. Reselling access and sharing accounts violates the terms of service of OpenAI and Anthropic. When a provider shuts down such an account, the key stops working along with any remaining balance you paid the seller.
• Data leak. All your requests and responses pass through the seller's server: prompts, code, and customer documents. There is no way to verify if they store them or who they show them to.
• Model spoofing. Instead of the stated model, the seller may respond with a cheaper one. The model field in the response is generated by their server, so it does not serve as proof.
• Opaque pricing. A "package of 100 thousand tokens" or "unlimited per month" is difficult to compare with official pricing, and "unlimited" usually implies hidden restrictions.

How to check a seller before paying:

1. Ask for the price per 1M input and output tokens and compare it with the official one from the table above. If GPT or Claude are significantly cheaper than the official price, it is likely a shared or stolen account.
2. Find out where the balance and usage are visible. A professional service provides a dashboard with a history, not a messenger message saying "40% remaining."
3. Ensure the key is issued personally to you and that you can delete it yourself. A login and password for someone else's account is not an API purchase, but a lease of access that can be revoked at any time.
4. Start with a minimum amount and run your real tasks on the key: compare the quality of responses and speed with the official version of the same model.
5. Do not send personal customer data, passwords, or proprietary code through such a key.

JoinGonka Gateway works differently: it does not resell OpenAI or Anthropic keys but provides access to open models on the Gonka network via your own keys. You create the key yourself in your dashboard and can delete it at any time, usage is visible by day, model, and individual key, and prices per 1M tokens are published on the website and in the API. Learn more about how the service works in the JoinGonka Gateway overview.

1. Registration. Open gate.joingonka.ai and create an account via email and password or sign in using Google, GitHub, or Discord. No bank card or crypto wallet is required for this.
2. Free tokens. Confirm your email — 3M tokens will be immediately credited to your balance. This is enough to connect the key to your application before your first payment.
3. Top-up. In the "Billing" section, click "Top up" and choose a method:
   • USDT — from $1 via the OxaPay payment service, 5% fee, lower on large amounts;
   • GNK — Gonka network token, zero fee;
   • WGNK from Ethereum — 1% fee, transfer is credited automatically;
   • bank card Visa or Mastercard — via a partner, from $10. All fees are visible before payment; the partner may ask to verify your identity.
   USDT and GNK are credited as soon as the transfer is confirmed on the network — usually in a few minutes; WGNK — usually in 15–20 minutes, card payments — in 5–30 minutes after the payment.
4. API key. In the "API keys" section, click "Create key." The key starts with jg- and is displayed only once — save it immediately.
5. Connection. In your application, specify the base URL https://gate.joingonka.ai/v1 and your key — this works for any OpenAI-formatted client. Tools with Anthropic formatting, such as Claude Code, require ANTHROPIC_BASE_URL=https://gate.joingonka.ai.

The list of available models and their identifiers is public without a key — insert the required one into the model field of your application:

curl https://gate.joingonka.ai/v1/models

Ready-to-use examples in Python, TypeScript, and curl are available in the API quickstart.

Consumption depends more on the task than on the model. Benchmarks for typical scenarios:

JoinGonka prices — $0.0069 per 1M input and $0.021 per 1M output tokens. In dollar terms: $1 on your balance is about 140M input or 48M output tokens. The 3M free tokens after registration are enough to connect a key and run your real tasks before your first payment.

How to keep costs under control. Create a separate key for each project: in the "Usage" section, consumption is visible by day, model, and key. Limit response length with the max_tokens parameter and avoid sending unnecessary context to the model — in agents, this consumes the majority of input tokens. If you sell access to your clients, issue them child keys with daily and monthly limits — they are created in the same "API keys" section.