A zero data retention LLM API

Every request Nivana sends to a model asks the provider to keep nothing and train on nothing. Here is exactly what that covers.

What happens to a request

• At the provider. Every call goes through Vercel AI Gateway with zero data retention and no prompt training requested, so only providers that honour both are used for it.
• At Nivana. API message content is not stored. Nivana keeps request metadata for routing and billing: your account id, the model, the task and difficulty, token counts, timing and cost.
• Long conversations. When a conversation is summarised to fit the context window, the summary is kept in memory briefly and handed back to you sealed; nothing is written to storage.
• Keys. Only a hash of each API key is stored.

Why it matters

Most teams can use frontier models only if prompts are not retained or used for training. Negotiating that with every lab separately is slow; with Nivana it is the default for every model the router can choose, including the ones it adds later.

Every model, one policy

The router picks from 14 models across Anthropic, OpenAI, Google, DeepSeek, xAI, Mistral, Moonshot and Meta (see the models), and the same retention rules apply whichever one answers. Read the full privacy policy and data handling in the docs, then get a key.