turboprivate - API Models
An OpenAI- and Anthropic-compatible model API on hardware we operate. Every session runs in an encrypted, ephemeral Room: hybrid X25519 + ML-KEM-768 key exchange, XChaCha20-Poly1305 transport, Ed25519-signed receipts. No prompt logging, no training on your data, nothing retained after the Room ends. Rooms are not hardware-attested today: the connection is sealed end to end, but a host operator could inspect live Room memory. We say so here, in the docs and in the API's trust state.
what it is
Your process creates the Room capability and establishes hybrid X25519 + ML-KEM-768 keys with the Room itself. The account service brokers and meters; it never receives a usable capability or model content.
The Room holds the model context only while inference is live. Ending the session destroys its keys and state. We keep account, balance and token counts—never the text.
Every answer carries an Ed25519-signed receipt, and the trust state of the Room (transit-only or hardware-verified) reaches your client before the first token. The SDKs and CLI expose both.
what stays where
| where | what it can see |
|---|---|
| Your process | Everything: plaintext prompts, answers and the Room keys. The SDK or CLI runs on your machine. |
| Account service | Your key, balance, model id and token counts. Never plaintext, never the usable Room capability. |
| The Room, while live | The model context and answer, in memory, for one session. Request-content logging is disabled in the pinned engine configuration. |
| After the Room ends | Nothing of the session content. Account, billing and limited operational metadata remain. |
pricing
Prepaid credits, no subscription. We bill measured input, cached-input and output tokens — never elapsed time — at the prices on the billing page.
model capacity
API registration is open. If the models you need are currently offline, leave an email for capacity updates. No newsletter, no third parties.
Create an API account or sign in to see your balance and create API keys. Registration is open; model capacity is shown in the console.