AI Reserve Developer Documentation

Trust & Disclosures

Provider Data Handling

The data-handling characteristics of each AI model provider available through the AI Reserve gateway — whether the provider trains on API inputs, how long it retains request data, where it processes requests, and its role under our Data Processing Addendum. These are the characteristics referenced by our Terms of Service, and they reflect each provider's enterprise API terms (not consumer-product terms, which often differ). This page may be updated from time to time as provider policies change.

Last reviewed: August 17, 2026

How to read this page

AI Reserve routes requests to the model you select. The table below lists each model provider we route to and summarizes, from that provider's own published enterprise API terms:

  • Trains on API inputs? — whether the provider uses API inputs or outputs to train or fine-tune its models under its current terms.
  • Retention — how long the provider retains prompts and completions after processing.
  • Processing region — where the provider states requests are processed and stored.
  • Role — the provider's role under our Data Processing Addendum. A Subprocessor processes data only on our documented instructions. A provider that processes data for its own purposes (for example, one whose terms permit training on inputs) is an Independent controller, not a subprocessor, and is identified as such.
Pending verification. Where a characteristic could not be confirmed from the provider's official documentation as of the last-reviewed date, it is marked pending verification — confirm before relying on it. We do not guess: an unmarked entry reflects the provider's published terms as reviewed on the date above.
Choosing a model chooses a provider. The characteristics below differ materially between providers — in particular, DeepSeek (direct API) and Moonshot AI process requests outside the United States and their terms permit use of inputs to improve their services. If your organization's policy restricts where data may be processed, ask us to pin your account's allowed models — pinning is enforced at the gateway.

Provider table

Provider Trains on API inputs? Retention Processing region Role
OpenAI (direct API) No, by default.1 Inputs/outputs retained up to 30 days for abuse monitoring, then deleted. Zero Data Retention available on approval.2 United States Subprocessor
Anthropic (direct API) No, under commercial terms.3 Inputs/outputs deleted within 30 days. Note: Anthropic's "Covered Models" (including Claude Fable 5, which we route) require 30-day retention and are not eligible for Zero Data Retention.4 United States Subprocessor
AWS Bedrock No. Inputs/outputs are not used to train foundation models and are not shared with the underlying model providers.5 Not stored by default; retention depends on the account's Bedrock data-retention mode (abuse-detection retention may apply in default mode; none = zero retention). Our account's configured mode: pending verification.6 United States (us-east-1) Subprocessor
Google Vertex AI No. Google states customer data is not used to train or fine-tune foundation models, and prompts/completions for partner models (Anthropic Claude) are not shared with the model's publisher.23 Not stored by default — Google states prompts and responses for Vertex generative models are processed in memory and not logged for training; abuse-monitoring and caching behavior is configurable at the project level. Our projects' configured mode: pending verification.23 United States (us-east5) Subprocessor
Google Gemini API (paid tier) No, on paid services.7 Prompts/responses logged up to 55 days solely for abuse monitoring, then deleted.8 Global — Google's terms state data may be stored transiently or cached in any country where Google or its agents maintain facilities.7 Subprocessor
xAI (Grok) No, without explicit permission.9 Requests/responses stored encrypted for 30 days for abuse auditing, then auto-deleted. Zero Data Retention available for enterprise accounts.9 United States Subprocessor
Perplexity (Sonar) No.10 Zero data retention on the Sonar API — prompts/responses deleted after processing; only billing metadata (token counts, timestamps) retained.10 United States (AWS, North America)11 Subprocessor
DeepSeek (direct API) Yes. DeepSeek's privacy policy states inputs are used to train and improve its models and services.12 No published fixed deletion window; retained under its privacy policy. pending verification China (People's Republic of China) — DeepSeek states it collects, processes, and stores data on servers in the PRC.12 Independent controller
Moonshot AI (Kimi) Yes. Moonshot's platform privacy policy states user content is used to train and refine its models.13 No published fixed deletion window; retained under its privacy policy. pending verification China (PRC-headquartered provider). Moonshot's international platform terms state its servers are located in Singapore; the endpoint serving our account is pending verification. Treat data sent to Kimi models as processed outside the United States under PRC-affiliated jurisdiction.13 Independent controller
Alibaba Cloud (Model Studio) No — Alibaba Cloud's Model Studio privacy notice states customer data is never used for model training.21 Model Studio stores data generated from model and application calls under its service terms; no published fixed deletion window. pending verification21 United States (Virginia) for Qwen chat models (dashscope-us endpoint); Singapore for Wan video generation and Qwen image generation (qwen-image-3.0 / qwen-image-3.0-pro) — the dashscope-intl endpoint is the only region serving wan3.0-video and the Qwen Image 3 family. Alibaba states request data is stored in the selected region. The operator is a PRC-headquartered group: the platform labels Alibaba-served models China-based and serves them only to organizations that have not opted out of PRC-affiliated models.21 Subprocessor
Mistral AI (La Plateforme) No — Mistral's Terms of Use for the paid API state "We do not use Your Prompts and/or Your Outputs to train Our Model(s)". The free Experiment tier trains by default; our account is the paid tier.22 Inputs/outputs retained 30 rolling days for abuse monitoring, then deleted. Zero Data Retention available on the API on request.22 European Union — Mistral is a France-headquartered provider and states La Plateforme runs on EU-based infrastructure by default; exact serving region for our account: pending verification Subprocessor
Together AI No — training use is opt-in only and not enabled.14 Prompts/outputs not stored by default (zero-data-retention posture); temporary caching may be used for performance.14 United States Subprocessor
Fireworks AI No, without explicit opt-in.15 Zero data retention by default — prompts/generations exist only in volatile memory for the duration of the request; not written to persistent storage.15 United States Subprocessor
DeepInfra No — DeepInfra states it does not train on data submitted to its APIs (its docs note an exception for Google/Anthropic models hosted there, which we do not route through DeepInfra).19 Not stored to disk — inputs/outputs exist only in memory while the request is processed and are deleted after; inputs/outputs may be kept for a limited period for debugging under its terms.19 United States — provider states US-based data centers (one Canadian site, Toronto, opened July 2026) Subprocessor
Novita AI No — Novita states it does not train on customer data; its privacy policy states personal information is not used for model training and customer-content processing is governed by customer agreements. API-content-specific terms: pending verification.20 No published fixed deletion window for API request content. pending verification United States (San Francisco HQ). Serverless fleet spans multiple international regions; serving region for our account: pending verification Subprocessor
Baseten No — Baseten states customer data submitted to its inference APIs is not used to train models; enterprise terms govern API inputs and outputs.24 Bounded and configurable under its terms — no training-use retention, and zero-data-retention options are available on enterprise plans. Exact retention window and configuration for our account: pending verification.24 United States (San Francisco HQ). Multi-cloud GPU fleet spans additional regions; serving region for our account: pending verification Subprocessor
OpenRouter No — OpenRouter does not train models; its own prompt logging is opt-in and off by default.16 OpenRouter stores request metadata only (no prompt/response content) by default. Retention at the downstream serving endpoint is governed by that endpoint's own policy; per-route endpoint policies for our account: pending verification.16 United States (routing layer); downstream serving endpoint varies by route. pending verification Subprocessor (routing layer; see note)
NVIDIA (hosted NIM API catalog)
Opt-in only — organization training consent required		 Locked by default for every organization. Under the published API-catalog trial terms, NVIDIA may use inputs/outputs to improve its products and services, including AI models17. The default Nemotron models (nemotron-3-super, nemotron-3-nano, nemotron-3-ultra) therefore route to DeepInfra under its no-training terms (see the DeepInfra row above) and never touch NVIDIA. Separate, clearly-labeled NVIDIA-served options (nemotron-3-super-nvidia, nemotron-3-nano-nvidia, nemotron-3-ultra-nvidia) exist only behind an organization administrator's explicit training consent (see "Organization consent gates" below) — between August 10 and August 14, 2026 no NVIDIA-served options existed at all, and no traffic reaches NVIDIA unless an administrator has opted the organization in. For organizations that opt in: content used to provide the service during the session; usage logged for security/fraud/abuse monitoring; inputs/outputs may be used to improve NVIDIA's products including AI models. No published fixed deletion window. United States Independent controller (only after organization training consent; no requests without opt-in)
Tensormesh (serverless inference) Training-use terms: pending verification. Tensormesh-served models are locked by default for every organization and require an organization administrator's explicit retention consent in the AI Reserve console (see "Organization consent gates" below). No published deletion window; retention terms pending verification and not currently disableable for our account. United States pending verification Subprocessor (only after organization consent; pending verification)
ByteDance (BytePlus ModelArk) Training-use terms under the international ModelArk service: pending verification. ByteDance-served models (Seedream image, Seedance video) are labeled "China-based" (PRC-headquartered operator) and are unavailable to organizations that have not enabled PRC-hosted models. No published deletion window for the international ModelArk API. pending verification Singapore (BytePlus international, ap-southeast); operator headquartered in the PRC. Subprocessor (only after organization consent; pending verification)
fal.ai (image/video generation) No — fal's API services terms state client content is not used to train or develop its products, except for models designated "Pending Enterprise Ready".18 Request payloads stored 30 days by default (opt-out available per request); generated media stored on fal's CDN per configured lifecycle.18 United States Subprocessor

Notes & clarifications

Aggregators and open-weight models

Together AI, Fireworks AI, DeepInfra, Novita AI, Baseten, and OpenRouter serve open-weight models published by other organizations (for example, Meta Llama, Mistral, Qwen, DeepSeek, Moonshot Kimi, NVIDIA Nemotron, and OpenAI's GPT-OSS family). When a model is served by a US aggregator, the aggregator's data-handling terms apply — the model's original publisher never receives the request. For example, deepseek-v4-flash and kimi-k2.6 are served by Fireworks AI in the United States under Fireworks' zero-retention terms, even though the underlying models were published by DeepSeek and Moonshot. Likewise, the NVIDIA Nemotron 3 family (nemotron-3-super, nemotron-3-nano, nemotron-3-ultra) is served by DeepInfra: since August 2026 NVIDIA receives none of those requests (earlier traffic used NVIDIA's hosted NIM catalog — see the NVIDIA row above). The DeepInfra- and Novita-served routes (deepseek-v4-flash-deepinfra, deepseek-v4-flash-novita, qwen-3.8-max-deepinfra, qwen-3.8-max-novita) work the same way: the serving provider's terms above apply, and the model's publisher (DeepSeek, Alibaba) never receives the request. Only the direct-API rows above (DeepSeek direct, Moonshot's own API for kimi-k3 and moonshot-v1-*, Alibaba's Model Studio API for qwen-3.8-max, Mistral's La Plateforme API for mistral-medium-3.5, codestral-2508, and ministral-14b) send data to those companies.

AWS Bedrock model routing

Models with a -bedrock suffix route through AWS Bedrock in the United States. AWS operates the model inside AWS infrastructure; the model's publisher (Anthropic, Meta, Mistral, DeepSeek) does not receive prompts or completions, except where a specific Bedrock model requires provider data sharing and the account has explicitly enabled that mode — we have not enabled provider data sharing.

Google Vertex AI model routing

Models with a -vertex suffix route through Google Vertex AI in the United States (us-east5). Google operates the model inside Google Cloud infrastructure as a data processor; the model's publisher (Anthropic) does not receive prompts or completions, and Google does not use them to train its models. Retention is configurable at the Google Cloud project level (see the table above).

Provider failover (same model, different host)

If the provider serving a selected model fails or is saturated, the gateway may retry the request through a disclosed failover chain instead of returning an error. Where the identical model weights are hosted by a second provider on this page — for example, Anthropic Claude and Meta Llama models on AWS Bedrock — that same-weights deployment is tried first: the response still comes from the exact model selected, processed by a different disclosed subprocessor. Cross-model degradation rungs (for example, a smaller model in the same family) come only after the same-weights option. Every response served by a fallback names the serving model in the x-aireserve-served-model response header — failover is disclosed per response, never silent. Failover never sends a request to a provider that is not on this page, and never to an Independent controller provider, unless a client administrator has explicitly configured such an entry in their own organization's fallback chain; chain entries hosted in the PRC are excluded for organizations that have opted out of PRC-hosted models. Some model names carry no failover chain at all (for example, -bedrock and -vertex routes and deepseek-chat/deepseek-reasoner, which promise a single serving provider) — those fail with the provider's real error instead.

Organization consent gates

AI Reserve's default posture is that gateway traffic reaches only providers documented above as not training on API inputs, with disclosed, bounded retention. Models served by providers that soften either guarantee are locked by default for every organization and stay locked until an organization administrator explicitly opts in from the AI Reserve console:

  • Training consent unlocks models served by providers whose terms permit training on API inputs with no opt-out (DeepSeek direct API, Moonshot AI, and NVIDIA's hosted NIM catalog — reachable only via the clearly-labeled -nvidia options re-added behind this gate on August 14, 2026; the default Nemotron routes stay on DeepInfra and never touch NVIDIA).
  • Data-retention consent unlocks models served by providers whose retention can neither be turned off nor verified yet (Tensormesh).
  • PRC-server consent unlocks models served by providers operating under PRC jurisdiction (DeepSeek direct, Moonshot, Alibaba Cloud Model Studio, ByteDance BytePlus ModelArk).

A model behind more than one gate (for example DeepSeek's own API: PRC jurisdiction and training-permitted terms) requires every applicable consent. Locked models appear greyed-out in the model catalog with the required consent named, are excluded from chat and automatic routing, and API requests to them are refused. Failover chains never carry a request onto a model whose consents the organization has not given.

Zero Data Retention

Several providers offer stricter Zero Data Retention (ZDR) arrangements on request (OpenAI, Anthropic, xAI, Together AI). The table reflects the standard enterprise default that applies to gateway traffic today. Where ZDR has not been separately confirmed for our account, assume the standard retention window shown. If your organization requires ZDR on a specific provider, contact us.

What AI Reserve itself retains

Independently of the providers above, the AI Reserve gateway logs request metadata (model, token counts, cost, latency, status, and user/key attribution) to operate billing and spend governance. By default, the platform also stores prompt and response content to power chat history, chat search, audit, and the optional usage-classification analytics described below.

Usage classification (optional, per organization). When enabled, conversation text is processed to categorize AI spend by business function for the analytics dashboard. To keep classifications reviewable and to measure and improve classifier quality, the platform keeps snapshots of classified conversation text: a bounded snippet of the triggering prompt while a classification awaits human review, and per-turn copies of the classified user and assistant text. These snapshots stay inside the platform, are never used to train foundation models, and are included in every purge and deletion described below. Enterprise administrators can turn classification off entirely, in which case no conversation text is processed or captured for analytics.

Zero content retention (optional). Enterprise administrators can turn off content storage for their entire organization from the admin console ("Store prompt & response content"). While disabled, the platform persists no prompt or response text and performs no content-based classification — only the billing metadata above is retained.

Deletion and retention windows. Previously stored content — prompt/response messages and the classification snapshots above — can be permanently deleted on demand from the admin console, and administrators can set an automatic retention window (for example, 90 days) after which stored content is deleted on a schedule. Without a window, content is retained until deleted. When a user account is deleted, that user's stored content is deleted with it; when an organization is deleted, all of its stored content is deleted. Billing metadata is retained in each case. See our Terms of Service and Data Processing Addendum for the platform's retention commitments.

Organization instructions and context. Enterprise administrators can set organization-wide instructions (behavior and policy text) and organization-level context (shared knowledge such as company facts and terminology) that the gateway adds to the start of every conversation for every user in their organization, across web chat and API-key traffic. Both are off by default and opt-in per organization; while enabled, that content is sent to the AI model providers with every request. When active, users see a notice in the chat screen and can read the full text there.

Sources

Provider policy documents reviewed on July 23, 2026; DeepInfra, Novita AI, and Alibaba Cloud Model Studio entries reviewed and added on August 4, 2026; NVIDIA routing retired and the Nemotron models moved to DeepInfra on August 10, 2026; Alibaba image-generation (Qwen Image 3, Singapore) coverage added on August 11, 2026; Google Vertex AI and Baseten entries added on August 12, 2026 (Baseten onboarding ahead of first traffic). Providers may update their terms; this page is reviewed periodically and updated when material changes are identified.

  1. OpenAI — Business data privacy: https://openai.com/business-data/
  2. OpenAI — Data controls in the OpenAI platform: https://developers.openai.com/api/docs/guides/your-data
  3. Anthropic — How long do you store my organization's data?: https://privacy.claude.com/en/articles/7996866
  4. Anthropic — API and data retention (Covered Models): https://platform.claude.com/docs/en/manage-claude/api-and-data-retention
  5. AWS — Amazon Bedrock security, privacy & responsible AI: https://aws.amazon.com/bedrock/security-privacy-responsible-ai/
  6. AWS — Amazon Bedrock data retention modes: https://docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html
  7. Google — Gemini API Additional Terms of Service: https://ai.google.dev/gemini-api/terms
  8. Google — Gemini API abuse monitoring: https://ai.google.dev/gemini-api/docs/usage-policies
  9. xAI — API security FAQ: https://docs.x.ai/developers/faq/security; Enterprise Terms of Service: https://x.ai/legal/terms-of-service-enterprise
  10. Perplexity — Privacy & Security (Sonar API): https://docs.perplexity.ai/docs/resources/privacy-security
  11. Perplexity — FAQ (hosting): https://docs.perplexity.ai/docs/resources/faq
  12. DeepSeek — Privacy Policy: https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html
  13. Moonshot AI — Kimi OpenPlatform Privacy Policy: https://platform.kimi.ai/docs/agreement/userprivacy
  14. Together AI — Privacy Policy: https://www.together.ai/privacy; Privacy and security docs: https://docs.together.ai/docs/privacy-and-security
  15. Fireworks AI — Zero Data Retention: https://docs.fireworks.ai/guides/security_compliance/data_handling; Privacy Policy: https://fireworks.ai/privacy-policy
  16. OpenRouter — Data collection: https://openrouter.ai/docs/guides/privacy/data-collection; Zero Data Retention controls: https://openrouter.ai/docs/guides/features/zdr
  17. NVIDIA — API Trial Terms of Service (§2–3): https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA API Trial Terms of Service.pdf
  18. fal — API Services Terms: https://fal.ai/legal/api-services; Data Retention & Storage: https://fal.ai/docs/documentation/model-apis/media-expiration
  19. DeepInfra — Data privacy: https://docs.deepinfra.com/account/data-privacy; Terms of Service: https://deepinfra.com/terms
  20. Novita AI — Privacy Policy: https://novita.ai/legal/privacy-policy; Trust Center: https://trust.novita.ai/
  21. Alibaba Cloud Model Studio — Security certifications and privacy notice: https://www.alibabacloud.com/help/en/model-studio/privacy-notice; Region and endpoint information: https://www.alibabacloud.com/help/en/model-studio/regions/
  22. Mistral AI — Help center on training use: https://help.mistral.ai/en/articles/323757-do-you-use-my-user-data-to-train-your-artificial-intelligence-models; Privacy and data controls: https://docs.mistral.ai/admin/monitor-comply/privacy-data-controls; Data Processing Addendum: https://legal.mistral.ai/terms/data-processing-addendum
  23. Google Cloud — Generative AI and data governance (Vertex AI): https://cloud.google.com/vertex-ai/generative-ai/docs/data-governance; Anthropic Claude models on Vertex AI: https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-claude
  24. Baseten — Trust Center: https://trust.baseten.co/; Privacy Policy: https://www.baseten.co/privacy-policy/