Data policy
Last updated November 7, 2026
Anix is operated by Geraldo Grell, a sole proprietor in New York, doing business as Anix ("Anix," "we," "us"). This page describes exactly what happens to the data you send to the Anix API.
In short
- Zero data retention: we never log, store or train on prompts or outputs.
- Request content exists only in server memory while your request runs.
- We keep request metadata for 30 days for billing and abuse prevention.
- All inference runs in the United States.
How a request is handled
- Your request reaches our API gateway over an encrypted (TLS) connection. The gateway checks your API key and forwards the request. It does not record the request body.
- Our inference server, on a US GPU provider, holds your request in memory while the model generates the response.
- The response streams back to you through the gateway as it is generated.
- When the request ends, its content is discarded from memory. Nothing is written to disk.
Prompt caching
To make repeated prompts faster and cheaper, the server keeps the model's internal state for recently processed prompt prefixes (the KV cache) in GPU memory. This state is never written to disk or logs, is separated by API key so one customer's requests can't reuse another's, and is evicted automatically as new requests arrive, usually within minutes.
What we keep
| Data | Why | How long |
|---|---|---|
| API key ID, model, token counts, cached-token count, time, duration and response status | Billing and usage reports | 30 days, then kept only as monthly totals |
| IP address the request came from | Abuse and fraud prevention | 30 days |
| Monthly usage totals and invoices | Accounting and tax | As long as the law requires, generally up to 7 years |
| Error logs, which never contain request content | Fixing problems | 30 days |
Training
We never use prompts, outputs or metadata to train or fine-tune any model, and we don't share them with anyone who does.
Human access
No one at Anix reads prompts or outputs. Because they aren't stored, there is nothing to read.
Requests through OpenRouter and other platforms
When you reach an Anix-served model through a platform such as OpenRouter, the platform sends the request to us and we handle it exactly as described here. The platform's own policies cover what it stores.
Subprocessors
These companies run parts of our service. Each handles data only to provide that service to us.
| Company | What it does for Anix | Data it handles | Location |
|---|---|---|---|
| Cloudflare | Website hosting, DNS, encryption and our API gateway | Requests in transit; connection data such as IP addresses | Global edge network |
| Beam | Primary GPU servers that run the models | Request content, in memory only, while a request runs | United States |
| RunPod | Backup GPU servers, used only if the primary is down | Request content, in memory only, while a request runs | United States |
| Stripe | Payments and invoices | Billing contact and payment details | United States |
| Google Workspace | Messages you send us | United States |
Security
- All connections use TLS encryption.
- API keys are stored only in hashed form.
- GPU servers accept traffic only from our gateway.
- Every provider account uses two-factor authentication, and only the operator has access.
If a security incident affects your data, we will tell you without undue delay after we confirm it.
Changes
If we ever change this policy in a way that reduces your protections, we will email customers at least 14 days before the change takes effect.
Contact
Questions about this page: geraldo@useanix.com