Inference overview

Hyperbolic Inference gives you API access to open-source language models without renting or managing GPUs. You send OpenAI-compatible chat completion requests, and you pay per token from your credit balance.

​

What’s available

​

Who can use inference

To send inference requests, you need:

• A Pro plan, or an Organization. Personal accounts on the Free Tier can’t use inference. To upgrade, make a one-time deposit of $5 or more. There’s no subscription. Click Upgrade to Pro for $5 under Settings > Plan, or upgrade from the Model Playground. API access is restored within 30 seconds of upgrading. Requests made with an Organization API key, or from the dashboard while you’re working in an Organization, don’t require a Pro plan. See Plan.
• An sk_live_ API key. Create one under Settings > API Keys. Keys issued in the older format aren’t accepted. See API keys.
• Credits. Your balance must be above zero. See Billing.

​

Available models

Pass one of these IDs in the model field. Prices are in USD per million tokens, as of October 21, 2026. To get the list through the API, see List models.

​

Context and output limits

Limits are in tokens, confirmed October 20–22, 2026. The context window covers your prompt and the generated output together. For glm-5.3-flash and qwen3.8-27b, a request whose prompt plus max_tokens doesn’t fit in the context window returns 400. The prompt isn’t truncated. qwen3.8-27b has no separate output limit, so its output can use whatever the prompt leaves of the 196,608-token window. Models and prices change over time. The Models page always shows the current list and prices. If you have custom pricing, the Models page shows your price.

​

Models and pricing

The Models page in the dashboard lists every model you can call. Each model’s page shows:

• The model ID to pass in the model field of your request.
• The price per million tokens.
• A Model Playground to try prompts in your browser.
• An API panel with ready-to-copy request code.

​

How tokens are billed

Prices are per million tokens and may differ for each part of a request:

• Input tokens: the tokens in your prompt, including the system prompt and conversation history.
• Cached input tokens: input tokens served from the model’s prompt cache. These are a subset of your input tokens, not an extra charge. When a model has a cached input price, they’re billed at that lower rate. Otherwise they’re billed at the input rate.
• Output tokens: the tokens the model generates.

You aren’t charged for requests that return an error. The rate that applies is the one in force when Hyperbolic receives your request.

​

Personal and Organization billing

• Personal API keys bill your personal balance.
• Organization API keys bill the Organization’s shared balance. When the Organization’s balance runs out, requests are denied until an admin adds funds. Organizations are billed standard platform prices. See Organizations.

​

Model Playground

Each model’s page includes a playground for testing prompts before you write code. You can set a system prompt, temperature, top-p, and max tokens, and the responses stream into the chat. Playground requests use the same API and are billed the same way as API requests, to the personal account or Organization you’re working in.

• Reasoning models show their thinking in a collapsed Thinking section above the answer. If a model uses up Max Tokens on thinking before it starts its answer, the playground tells you to raise Max Tokens.
• Free Tier accounts see a notice that the model needs a Pro account. When you send a message, an Upgrade to Pro dialog opens with a $5 payment form. After the upgrade goes through, the playground sends your message.
• Organizations with no funds see an Add funds prompt when you send a message.
• Errors show the message from the API followed by its code, for example (insufficient_balance). See Errors.

​

Moving from the previous inference API

If you used Hyperbolic’s previous serverless inference API, update your integration: Requests to https://api.hyperbolic.xyz/v1 now return 404 with a message that the endpoint was decommissioned. Model IDs have also changed, so use an ID from Available models.

​

Next steps