The buyer-facing summary of how Pareta handles the requests
your application sends through the model ID auto: where they are
processed, who else processes them, how long they are kept, and what Pareta
does not do with them. Every fact on this page is taken from the
Privacy Policy (last updated 2026-08-01) and the
Terms of Service (version 2026-08-17), which govern;
nothing here extends them.
Pareta matches each request to the specialist that holds the workload’s quality bar, or to a frontier model when none does. That means your inputs are processed by:
A third-party AI provider may process a request in three ways: real-time verification, when a specialist result is checked before the response is served; frontier serving, when routing selects a frontier model to produce the response; and ongoing quality monitoring, when samples of completed traffic are re-run and scored after serving to measure and maintain answer quality.
The quality gate applied to a specialist's result before the response is served. It runs at request time and may use a third-party model as a grader.
A frontier model (OpenAI, Anthropic, or Google) produces the response when needed to meet the workload's quality bar — when the specialist path does not clear verification, or when no Pareta specialist holds the bar for that workload.
After serving, samples of completed traffic are re-run and graded on a rolling basis, including by third-party models acting as graders, to check that answers hold the quality bar. Human review is limited to cases the monitor flags as quality exceptions; staff access to stored request content is recorded in your organization's audit log.
Verification is not a frontier call on every request: output parsing and schema checks run on every specialist response; model-graded checks run on tasks whose measured quality calls for them. A frontier-served response is returned as served. Verification is a quality gate, not a guarantee that an AI output can never be wrong — AI model outputs are probabilistic and can be wrong; you are responsible for evaluating outputs before relying on them, and for human review where outputs affect health, legal, financial, or safety decisions.
What Pareta keeps, why, and for how long.
| Data | Why it is kept | Kept for |
|---|---|---|
| Prompts and responses | Service operation — debugging failures and investigating abuse — and quality monitoring. Applies to API traffic and playground use, and includes any documents or images in the request. | Up to 30 days, then deleted automatically. |
| Benchmark datasets and evaluation results | So you can re-run and compare over time. Deleting a dataset deletes its items, uploaded files, and run results. | Until you delete them. |
| Usage and billing metadata | Billing integrity, audit, and service analytics. Timestamps, token counts, latency, status, cost, and the routing decisions taken — without prompt or response content. | Indefinitely. |
| Organization audit log | Logins, key creation, admin actions, and any staff access to stored request content. | Kept. |
| Account data | To operate your account. Email address, name, a hash of your password, and the Terms of Service version you accepted and when. | For as long as your account exists. |
Pareta does not use your inputs or outputs to train models, and does not sell them.
Pareta operates primarily from the United States (us-central1). If you use the Service from elsewhere, your data is transferred to and processed in the US.
Pareta-hosted models run on Google Cloud and Modal, primarily in the United States. A regional-processing option is not offered today — see §8.
The subprocessor table from the Privacy Policy, §3, reproduced in full. The first five process request content — as infrastructure, inference, or routed and sampled requests; the remaining five do not.
| Provider | Purpose | Sees request content? |
|---|---|---|
| Google Cloud | Hosting, storage, databases | Yes (infrastructure) |
| Modal | GPU model serving | Yes (inference) |
| OpenAI | Frontier serving + quality grading | Yes (routed/sampled requests) |
| Anthropic | Quality grading; frontier serving | Yes (routed/sampled requests) |
| Google AI (Gemini) | Quality grading; evaluation runs | Yes (routed/sampled requests) |
| Stripe | Payments | No (billing data only) |
| Resend | Transactional email | No (email address only) |
| Cloudflare | Bot protection on signup | No |
| Pydantic Logfire | Operational telemetry | No (metadata only — no prompt content) |
| Slack | Support channels (opted-in orgs) | No (what you post there) |
Source: Privacy Policy §3. The policy’s table governs; this page is updated when it changes.
What the Privacy Policy and Terms state about transport, credentials, and API keys.
Access to stored request content by Pareta staff is restricted and recorded in your organization’s audit log.
Not offered today. There is no opt-out from the 30-day retention, and no option to process data outside the United States.
If your use case requires zero retention or regional processing, contact support@pareta.ai before sending production traffic.
No system is perfectly secure. Report suspected vulnerabilities to support@pareta.ai. Use the same address if you believe an account or API key is compromised.
Pareta Inc. · support@pareta.ai · Full documents: Privacy Policy (last updated 2026-08-01) · Terms of Service (version 2026-08-17).
Evaluate Pareta on representative examples before sending
production traffic — an evaluation set is 5–50 items, and the datasets you
upload stay until you delete them. Use the OpenAI client already in your
application and set the model ID to auto.