Skip to main content
How NoMask processes requests, records usage, and handles optional caches.

​
Content privacy

Billing records and operational analytics never contain:
  • Prompts, messages, and system prompts
  • Model responses and reasoning content
  • Images you send to vision models
  • Video files and inline video data
These records also exclude raw API keys and authorization headers. Request content is processed to fulfill your API call, not used for training by NoMask. Optional retry cache. When you send an Idempotency-Key on a non-streaming completion, its successful response payload is cached for 24 hours so matching retries can receive the same response without another charge. That payload can include response text, reasoning and tool calls. This cache is separate from usage analytics and does not apply to streaming requests. See Idempotency. Model caching. Model processing may temporarily retain reusable input. Retention varies by model; NoMask does not promise zero retention for all model processing. See Prompt Caching. Private
Your prompts and replies aren’t stored or used for training under Zero Data Retention (ZDR). Anonymized
Anonymized prompts may be retained and used for training.

​
What is stored

A billing metadata record is kept for each settled completion request. It contains: the request id, model id, token counts, the USD cost and rate snapshot, whether it streamed, latency, and the account and key used. GET /usage exposes your settled-request history; its response fields are listed in the endpoint reference. For newly settled completions, separate operational records track the completion or stream outcome, timing, and feature flags or counts for image and video inputs, tools, reasoning, response format and the NoMask system prompt. These records use the billing request id and contain no prompt or output content, source URLs, or API key secrets. They do not record rejected requests or unbilled failures, and are not additional fields in the public GET /usage response. Older requests may not have these operational metrics. Requests rejected before generation are not billed. After a stream starts delivering model output, consumed tokens may still be billed if it later fails or disconnects. A recorded outcome is not, by itself, a charge.

​
Anonymized inference

Inference requests carry no identity: no name, no email, no account id, no IP is attached to the content that reaches a model. Your prompts are indistinguishable from anyone else’s.

​
Account deletion

Deleting your NoMask account permanently removes every API key, the full usage history, all funding invoices, and the remaining balance. This is immediate and irreversible.