Negbeta

Inference API Privacy Policy

Effective: 28 October 2026

1. About this policy

This policy covers the Negbeta inference API at inference.negbeta.com (the “API”), which serves AI language models to marketplaces such as OpenRouter that resell our API and, through them, to their users. It is operated by Negbeta Limited (“Negbeta”, “we”, “us”), a company registered in England and Wales (company number 14318349), registered office 71-75 Shelton Street, Covent Garden, London WC2H 9JQ, United Kingdom.

The privacy policy for our corporate website is separate. Exactly which request data we keep, field by field, is listed in our API data policy.

For the content of requests we act on behalf of the marketplace that sends the request, which decides what is sent to us. For the operational metadata described below, Negbeta Limited is the data controller. We are registered with the UK Information Commission, the data protection regulator, under registration reference ZC282184.

2. The short version

  • We do not train on your data. Prompts, images and completions are never used to train or fine-tune any model, by us or anyone else.
  • We do not store prompts or completions. Prompts, images and model output are held in memory only while your request is processed, and are never written to our logs or database.
  • We keep per-request metadata for 180 days. Token counts, timings, sampling parameters and similar operational data, then it is deleted automatically. Raw IP addresses and user agents are deleted after 30 days.
  • Inference runs in the European Union, and for some models in the United States. Most models are served from servers we rent from Nebius in the European Union (and, where a marketplace asks for it, in the United Kingdom). Some models are served from GPU servers we rent through GPU cloud and marketplace providers in the European Union and the United States. A marketplace that presents our service as hosted in the European Union keeps its traffic in the European Union. This page lists every location we use.

3. How requests reach us

We do not offer accounts to the public. Requests reach the API from marketplaces such as OpenRouter that resell our API, each authenticated with an API key that we issue to that marketplace, over an encrypted connection through Cloudflare. We do not receive your name, email address or marketplace account details. The network address we see is normally the marketplace’s, not yours.

4. What we process and why

Request content

Your messages, system prompts, images, tool definitions and tool results, and the model’s output (including reasoning text and tool calls), are processed only to produce the response to your request. While the request runs they exist in the memory of our servers. To make repeated prompts faster, the server keeps a cache of recently processed prompt beginnings, in the model’s internal form, in temporary storage on the same server (today its GPU memory and main memory; in future also its local disk). That cache is stored securely, is used only for prompt caching and for nothing else, is overwritten as new requests arrive, and today is cleared whenever the server restarts. If a request links to an image by URL, our server downloads the image to process the request and does not keep it.

Operational metadata

For every request we record operational metadata: identifiers, timings, token counts, sampling parameters, message and tool-call counts, a short sanitised error message when something fails, and a one-way hash of the API key. We record the IP address of the connecting client, its country and user agent. For API traffic the client is the marketplace’s servers, not end users. Raw IP addresses and user agents are deleted after 30 days; other request metadata after 180 days. The full list is in the API data policy. We use it to:

  • reconcile usage and billing with the marketplaces that resell our API;
  • run the service: capacity planning, routing, monitoring and fixing faults;
  • measure and improve quality, for example the share of tool calls that are valid;
  • investigate abuse of, or attacks on, the service.

Our lawful basis for this metadata is our legitimate interest in operating, securing and billing the service. We do not use it for advertising or profiling, and we do not sell it.

5. Where data is processed

Most of our models are served from servers we rent from Nebius in the European Union, where our metadata database also runs. Where a marketplace asks for UK capacity, inference for that marketplace runs in Nebius’s UK data centre instead. Some models are served from GPU servers we rent through GPU cloud and marketplace providers (currently RunPod and Vast.ai) in the European Union and the United States; the locations can differ by model. Request metadata is stored only in the European Union. Our staff run the service remotely from the United Kingdom, which the European Union recognises as providing adequate protection for personal data. We may add capacity in other locations, in the European Union or elsewhere. Each marketplace’s traffic is served from the locations agreed with that marketplace: where a marketplace presents our service as hosted in the European Union (for example with an EU label or region), its traffic stays in the European Union unless that marketplace agrees otherwise; for other marketplaces we may serve from any of our locations and move between them. This page always lists the locations we use, and for a location outside the European Union and the United Kingdom (today, the United States, for some models) we use the safeguards that data protection law requires: today, standard contractual clauses. Requests travel to those servers through Cloudflare’s network, which may carry them through Cloudflare locations outside the European Union on the way. How a marketplace handles your data before it reaches us is covered by that marketplace’s own privacy policy.

6. Service providers

We use these service providers that handle request data on our behalf:

  • Nebius: the cloud servers and GPUs that run most of the models and store the metadata, currently in the European Union (and in the United Kingdom where a marketplace asks for UK capacity).
  • GPU cloud and marketplace providers (currently RunPod, Inc. and Vast.ai, Inc.): GPU servers that run some of our models, currently in the European Union and the United States. On Vast.ai, a marketplace, each server is run by an independent host. Request content is in these servers’ memory only while a request runs, plus the temporary prompt cache described above; our metadata database does not run on them. Both are US companies, so we use them under their data processing agreements and standard contractual clauses, wherever the server is.
  • Cloudflare, Inc.: the network edge in front of the API. Cloudflare terminates the encrypted connection, so request content passes through its network in transit; it does not store it for us. Cloudflare also answers requests for the model list and the home page itself, and keeps a record of each such request for us (time, address requested, raw IP address, country, network operator and user agent of the client).

We share no other data with third parties, except where the law requires it (for example a court order).

How we process request data on behalf of the marketplaces that resell our API, including the full list of our subprocessors, is set out in our Data Processing Agreement.

7. How long we keep it

  • Request content: not kept after the response is complete (the temporary prompt cache described above aside).
  • Per-request metadata: 180 days in our database, then deleted automatically. Raw IP addresses and user agents: 30 days. Temporary local copies made when the database is unavailable are described in the API data policy.
  • Records of requests for the model list and the home page, which are not API calls (including raw IP address, country, network operator and user agent): three months, kept by Cloudflare, then deleted automatically.
  • Server performance metrics (GPU load, queue lengths and similar, with no request content): 90 days.
  • Service logs on our servers (start-up and error events with request IDs, never request content): deleted after 14 days. Our metadata database’s own log of start-up and error events is limited by size, not by age.
  • Our database applies its time limits (180, 90 and 30 days above) while it runs. If the server is stopped (for example between test periods), records that pass their limit during the stop are deleted within a day after it starts again.

8. Security

The API accepts only encrypted connections and requires an API key; our model servers are not directly reachable from the internet. The metadata database is not exposed to the internet and is reachable only from our own servers; we run our monitoring dashboards on our own server too, reachable only over our private network. Access is limited to Negbeta staff who run the service. As with any rented server, the provider and the host operating a GPU server we rent through RunPod or Vast.ai also have physical and administrative access to it.

9. Your rights

Under the UK GDPR and the Data Protection Act 2018, and, where it applies (for example to personal data of people in the European Union), the EU General Data Protection Regulation, you have rights to access, correct, delete, restrict or object to the processing of your personal data. You can complain to the UK Information Commission, which replaced the Information Commissioner’s Office (ico.org.uk), or, in the European Union, to the data protection authority of the country where you live or work or where you believe the infringement took place. We store no names or email addresses, and the network addresses we store are normally those of the marketplace that sent the request, so we usually cannot tell which metadata relates to you. If you contact us with a request ID returned by the API, we will find and handle the matching record. Requests about your account with a marketplace (for example OpenRouter) should go to that marketplace.

10. Children

The API is a developer service and is not directed at children.

11. Changes

If we change how we handle request data, we will update this page and the “Effective” date before the change takes effect.

12. Contact

Email [email protected], or write to Negbeta Limited, 71-75 Shelton Street, Covent Garden, London WC2H 9JQ, United Kingdom.