TAO-1 API: privacy and data retention
Effective: October 14, 2026 (03:12 UTC) · Operator: Tao Synergies Inc. · Contact: support@tao1.ai
In short
- For requests made through the API, we do not keep the text of your prompts or of TAO-1's answers. We keep only per-request details: request id, model, token counts, timing, status and cost, plus a short fingerprint of the API key.
- We never use API requests to improve TAO-1, and we do not train or fine-tune any model on your prompts or answers. There is no opt-in.
- xagram sends your request to the AI provider that answers it. Providers handle data under their own policies, and some may keep it or use it to improve their models. Don’t share anything you wouldn’t want a third-party provider to see. The providers are listed in section 4.
- We do not sell prompts and do not share them for advertising.
1. What this covers
The TAO-1 API (/v1/chat/completions and related endpoints), whether you call it directly with a TAO-1 key or through a router, marketplace or reseller that forwards your request to us. If a router calls us for you, the router's own terms also apply to what it keeps. Our apps (tao1.ai and jtao.ai: chat and Build) have their own privacy page; the 100-day retention described there applies to app traffic only.
2. What we keep for each API request
- Request details: time, request id, model route, token counts, latency, status, cost, safety verdicts (labels only) and a short fingerprint of the API key (not the key itself). No prompt or answer text. Kept for billing, abuse control and reliability.
- Prompt and answer text: not stored. The request is processed to produce the answer and then dropped.
- Billing records: top-ups, holds and usage by key. They contain no request text. Kept for legal and accounting reasons.
- Our application logs do not contain request bodies. Our hosting providers keep their own platform logs (IP address, path, status code) for a short time.
Before October 14, 2026 (03:12 UTC). Requests made earlier were stored for up to 100 days and are deleted on that schedule. They may remain in encrypted backups for up to 8 weeks after deletion (section 5).
3. Training and improvement
- No model training. We do not use customer prompts or answers to train or fine-tune a model. Benchmarks and evaluations use public datasets.
- No learning from API traffic. API requests are never added to our improvement data. The improvement setting in our apps does not apply to the API. Request flags such as
metadata.no_improveandmetadata.learn_opt_inare still accepted and change nothing: the answer is already no.
4. Who processes your prompts
To answer a request we send its content to the provider that serves it. Today (from our last seven days of traffic):
- Anthropic (Claude Sonnet and Opus), directly and through OpenRouter on Amazon Bedrock and Google Vertex.
- Google (Gemini), directly through Google AI Studio and through OpenRouter on Google Vertex.
- Baseten (DeepSeek models), our largest share of requests.
- Engy (api.engy.ai): open models such as Kimi, DeepSeek and Qwen. Engy handles data under its own policy, which we do not control. It keeps usage details such as time, model, tokens, cost and latency.
- OpenRouter, a route that forwards to the underlying model vendors. In the last week these included Mistral, Meta, Together, Fireworks, DeepInfra, Z.AI, MiniMax, Parasail, StreamLake, CoreWeave, Amazon Bedrock and Google, and more may be added (Baidu may be among them). The underlying model vendors handle data under their own policies, and some may keep it or use it to improve their models.
- SayGM, a model gateway that some requests use.
- Typesafe.ai reads each request so that it can be sorted into the right kind of task.
- RunPod runs our own endpoints for image generation, speech to text, video and an image safety check.
- Web search (when a request needs current information): Desearch, with SerpAPI and Serper as fallbacks, receive a search query derived from the request, not the whole conversation. Desearch also reads pages.
- Public data sources (weather, prices, filings and similar) receive only the name or ticker needed.
We update this list when it changes.
5. Backups and security
Data is stored in managed PostgreSQL databases on our hosting providers. Nightly backups are encrypted and deleted after up to 8 weeks (14 daily and 8 weekly copies). Anything deleted from the live database may therefore remain in an encrypted backup for up to 8 weeks, and is restored only for disaster recovery. Access is limited to the operating team and service credentials. API keys are never stored in request records, only a short fingerprint.
6. Deletion and your rights
- Because API request text is not stored, there is no prompt or answer text for us to delete. To delete request details or billing records tied to your account, or to delete your account, write to support@tao1.ai; billing numbers needed for accounting remain.
- Router users: ask the router to forward the request to us, or write to us with your account or request id.
- Where the law gives you them (EU, UK, Switzerland, California), you may ask for access, correction, deletion and portability.
7. Changes and contact
We will update this page and its effective date when our practices change. Questions and requests: support@tao1.ai. See also our Terms of Service.