Zero data retention
Zero data retention is the default for this API. Prompts, outputs and request-specific intermediate content are not stored, unless you turn on Contribute to Fortytwo models or use a free model.- This includes traces that contain content, reconstructed prompts and embeddings of your content.
- Fortytwo does not call tools and does not store tool definitions, arguments or results.
store: trueis rejected with400 zdr_storage_not_supported.
No training on customer content
Customer content is not used to train models by default. Requests to free models are the exception. This applies to every model that serves the request: base models, swarm, draft, decoder, ranking and reward models, and models used for inference upsampling. A content-sharing programme requires a separate express opt-in. Before the opt-in, the programme states what is covered, the purpose, the period, who has access and how to withdraw. Accepting the Terms or a cookie banner is not an opt-in.Contribute to Fortytwo models
Contribute to Fortytwo models in account settings is the opt-in programme. It is off by default and can be turned off at any time.
While the setting is on, do not include passwords, API keys or other secrets in requests.
See Privacy Policy.
Free models
A free model has an ID ending in:free. Every request to a free model is collected and used to improve Fortytwo products and services, including evaluating and training models, and to keep them secure, whether or not Contribute to Fortytwo models is on. Staff who handle security and abuse can also access it. The rest follows the Contribute to Fortytwo models table. To stop sharing, stop using free models.
Do not send confidential information, secrets or personal data to a free model.
Abuse checks
Automated safety and abuse checks can run on requests while they are processed. If a check flags a request as a probable serious violation of the Terms of Service, the flagged content and related identifiers can be kept for up to 30 days. They are used only to investigate, to enforce the Terms or to meet a legal duty, and only designated staff can access them. A documented investigation, dispute or legal requirement can extend the period. Flagged content is never used for training. This is the only content kept without the customer’s choice.Prompt caching
A prefix of the prompt can be served from a short-lived cache. This sets the cached-input price — see Pricing.
A cache miss is charged at the full input price and gives the same result.
usage.prompt_tokens_details.cached_tokens shows how many prompt tokens came from the cache. See Privacy Policy.
Where inference runs
Inference runs on Fortytwo-controlled infrastructure in the United States. Requests are not routed to independent public node operators. The provider register lists every provider that processes data for Fortytwo — see Privacy Policy.Retention
See Privacy Policy.