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: October 20, 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. Where a characteristic instead comes from a written confirmation by the provider or its partner organization rather than published terms, the entry says so inline and cites the confirmation in Sources.
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 Depends on the model. Our account's Bedrock data-retention mode in us-east-1 (the Region every Bedrock request is made from) is aws_review, set October 8, 2026 because Anthropic requires it for Claude Fable 5 and Fable 5.1 on Bedrock. For those two models, prompts and completions are retained inside the AWS boundary for up to 30 days and may be reviewed by AWS for abuse detection; they are not shared with Anthropic. Models whose Bedrock terms permit zero retention — every Claude model released before Fable 5, including Claude Opus 4.8 — are not retained regardless of the account mode. Other Bedrock models follow each model's default: AWS may retain inputs/outputs for abuse detection; the model provider does not receive them. For cross-region requests, any retained data is stored in the AWS Region that processed the request.6 Global — AWS commercial regions via Bedrock global cross-region inference, requested from us-east-1 (-bedrock models); United States only for -bedrock-us models 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 Documents saved to the web document library are additionally held in a per-organization Gemini File Search store, with no expiry until deleted — see saved documents below. 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 The platform can also route text-to-speech to xAI (input text processed like any request) and supports customer-created custom voices — a capability that is organization-gated and off by default for every organization. If an organization's administrator opts in, reference audio uploaded to create a custom voice (a recording of a person's voice) is transmitted to xAI, which builds and stores the resulting voice model keyed to the voice id until the organization deletes it; AI Reserve itself stores no audio, only the ownership and consent record. Creation requires an explicit consent attestation for the voice being cloned. United States Subprocessor
ElevenLabs (text-to-speech)
Data-retention consent required		 Not under the account settings we route under. ElevenLabs' Privacy Policy states it uses personal information to "research, build, train and improve" its AI models, with an account-level opt-out ("Data use" in the account's Terms and Privacy settings) that applies from the moment it is set; AI Reserve sets that opt-out on its ElevenLabs account before any customer request is served.31 Verification of the opt-out on our account pending — listed ahead of deployment. By default ElevenLabs retains the text sent for synthesis and the generated audio in the account's request history "to enhance services, troubleshoot issues, and ensure the security" of its systems, with no published fixed deletion window; history entries are deletable from the account. Zero Retention Mode (immediate deletion of TTS text and audio, enable_logging=false) is offered to enterprise customers only and is not available on our account, so the eleven-* models sit behind the organization's data-retention consent gate.31 AI Reserve itself stores no audio — the usage row records the character count and the voice id only. United States by default (ElevenLabs states customer data is hosted and stored in the U.S.; processing may also occur by its international affiliates and subprocessors for support and content-moderation purposes).31 Subprocessor for text-to-speech once the training opt-out is set; serves only to organizations that have opted in to the data-retention gate.
Tako (live data and web search)
Data-retention consent required		 Not for paid-plan queries. Tako's Privacy Policy states it uses Queries "to provide the Services and to evaluate and improve them, including by training the search and ranking models" but that this "doesn't apply to Queries from customers on a paid plan"; AI Reserve's Tako account is a paid, pay-as-you-go account. The LLM providers Tako forwards Queries to operate under zero-data-retention agreements and do not train on them.32 Verification on our account pending — listed ahead of deployment. Tako retains the search Queries submitted through its API and processes them to fulfil the request; no published fixed deletion window. Zero Data Retention (Tako not retaining a customer's Queries) is arranged only under an enterprise agreement, which our account is not on, so the tako-search and tako-search-deep models sit behind the organization's data-retention consent gate.32 Only the query text is sent — never the surrounding conversation; AI Reserve's usage row records the query length, effort, result counts and Tako's request id, not the query itself. United States (Tako Inc.; processing region per Tako's own statements).32 Subprocessor for live search; serves only to organizations that have opted in to the data-retention gate.
Cloudflare (Workers AI — Clef decision models, Clairic extraction model) No. Cloudflare's Workers AI data-usage terms state that inputs and outputs are Customer Content and that Cloudflare "does not use your Customer Content to (1) train any AI models made available on Workers AI or (2) improve any Cloudflare or third-party services, and would not do so unless we received your explicit consent"; no such consent is given.34 Not stored on this path. Cloudflare states Workers AI Customer Content "may be stored by Cloudflare if you specifically use a storage service (e.g., R2, KV, DO, Vectorize, etc.) in conjunction with Workers AI"; the clef and clef-flash decision models and the clairic extraction model are called through the plain inference API with no Cloudflare storage product, AI Gateway or logging feature attached, and Cloudflare does not make Customer Content available to any other customer.34 The request state, typed questions or fields and any embedded images are sent and the probabilities or extracted values returned; AI Reserve's usage row records the question or field count and types, image count, token count and Cloudflare's request id, not the state itself. No consent gate applies. Cloudflare's global network (Cloudflare, Inc., United States); the model runs at a Cloudflare location near the gateway's U.S. region, and no processing-region pin is offered on the self-serve plan.34 Subprocessor for decision models (POST /v1/decisions) and the extraction model (POST /v1/extractions); opt-in by construction — nothing is sent to Cloudflare unless an organization calls a route.
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), except qwen-3-coder-next, which is served from Singapore (dashscope-intl endpoint — the only region offering that model, effective September 1, 2026); 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
MuleRouter (CarrotHub / SmartStudio ecosystem) No — MuleRouter does not retain request data for training, confirmed in writing by the Alibaba Cloud International partner team (August 26, 2026; written vendor confirmation, not yet published terms).26 MuleRouter's documentation states generated outputs belong entirely to the user. No published fixed deletion window. Deletable on request: we hold the right to delete tasks run on MuleRouter, confirmed in writing by the Alibaba Cloud International partner team (August 26, 2026; written vendor confirmation, not yet published terms).26 Singapore — confirmed in writing by the Alibaba Cloud International partner team (August 25, 2026).26 The operator is an Alibaba Cloud–partnered entity serving the PRC-developed W3.0 (Wan 3.0) video family without upstream content moderation (including, since October 15, 2026, Wan 3.0 video extension driven through the family's reference-video mode; the platform reads only the source clip's header to measure its length before submitting, and stores none of the clip). The platform labels these models China-based and serves them only to organizations that have not opted out of PRC-affiliated models; additionally they are allowlist-restricted, adult-content-gated, and third-party-policy-gated, and never publicly listed. Subprocessor
WaveSpeedAI PTE. LTD. Serving since September 1, 2026 (Wan 2.7 video extension; the launch gate was enabled by a reviewed change), with Wan 2.7 image editing added September 2026 and Wan 3.0 video extension added October 2026 on the same account and launch gate. Training use for inference inputs: pending verification. WaveSpeedAI's privacy policy scopes "Training Data" to content customers explicitly upload to train models, which our integration never does. WaveSpeedAI's documentation states generated media URLs generally expire within 7 days; retention/deletion terms for request inputs pending verification. Singapore-incorporated operator (WaveSpeedAI PTE. LTD., per its published privacy policy); serving region and upstream affiliations pending verification — conservatively labeled China-affiliated until verified. When enabled, the models are allowlist-restricted, adult-content-gated, third-party-policy-gated, served only to organizations that have not opted out of PRC-affiliated models, additionally behind the unverified-retention consent gate while the retention terms above are pending, and never publicly listed. Subprocessor
monid.ai (data-endpoint marketplace) Not published — monid.ai publishes no terms on training use of request inputs pending verification. monid is a broker: the inputs a customer sends to /v1/monid/run are forwarded by monid to the third-party data endpoint the customer selected (for example an Apify actor or a People Data Labs enrichment call), whose own terms then apply. Not used for AI Reserve model training in any case (see training). Not published — monid.ai publishes no retention window for run inputs or outputs pending verification. Run outputs are returned to the caller and are not rehosted by AI Reserve. Once a run completes, the stored request is reduced to its billing skeleton (endpoint identity, price, units) — the inputs the customer supplied are not kept — but the run’s result body, including the data the endpoint returned, stays on the run record so it can be re-read at /v1/monid/runs/{id}, and is deleted with the organization. Not published pending verification. Requests carry the endpoint inputs the customer supplied, on AI Reserve's single monid workspace; the downstream data provider varies per run and is named in every discover / inspect / run response. Subprocessor (data-endpoint runs only; never receives chat, completion, or document content)
TypeSafe AI (Jev) No. TypeSafe states that customer data does not influence model training; customization happens only through request parameters.28 Not published — TypeSafe publishes no retention window for request state pending verification. Used for two distinct purposes: native Auto model selection and optional per-request usage classification. Auto model selection sends a bounded request excerpt (up to 2,000 characters), organization routing instructions (up to 4,000 characters), routing preferences, and eligible model identifiers with available pricing/capability metadata; it does not send the whole conversation, system prompt, tool outputs, or completions. The separate opt-in usage classifier receives the excerpt described there. Jev returns typed choices and probabilities, not generated text. AI Reserve does not persist the router state or classification excerpt in router/classifier telemetry. Not published pending verification Subprocessor (native Auto routing and separate opt-in usage classification; bounded excerpts, not full conversations)
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 an organization-level opt-in only and is not enabled; under Together's Terms of Service (§2, Zero Data Retention) content is not used for model training or product improvement.14 Zero data retention — Together's Terms of Service (§2) commit that, with the organization's "store prompts and model responses" setting off (our configuration), prompts and outputs are not stored, retained, or used beyond the time required to process the request; a short-lived serverless prompt cache may hold content briefly for performance.14 United States Subprocessor
Fireworks AI No, without explicit opt-in (for example, Fireworks' FireOptimizer feature) — confirmed in writing by Fireworks AI (September 8, 2026; written vendor confirmation corroborating its published terms).15 Zero data retention by default — prompts/generations exist only in volatile memory for the duration of the request (a few minutes longer with prompt caching) and are never written to persistent storage; not shared with external model providers. Confirmed in writing by Fireworks AI (September 8, 2026; written vendor confirmation corroborating its published terms).15 United States Subprocessor
DeepInfra No — DeepInfra's Terms of Service (§7(b)) commit not to use Customer Data to train, fine-tune, or otherwise improve any model (its docs note an exception for Google/Anthropic models hosted there, which we do not route through DeepInfra).19 Zero data retention — DeepInfra's Terms of Service (§7(b), "Zero Data Retention") commit that Customer Data is not retained, stored, or logged beyond the period strictly necessary to process and return the request; inputs/outputs exist only in memory while the request is processed. Content may be retained for debugging only at the customer's written request or with prior written authorization, and is deleted within 30 days of the issue being resolved. §7(a) makes this commitment control over its privacy policy and documentation.19 United States — provider states US-based data centers (one Canadian site, Toronto, opened July 2026) Subprocessor
Novita AI No — Novita's Terms of Service (§10.2, "Zero Data Retention") commit that, by default, customer Content is not used to train its models or improve its services; its privacy policy separately states personal information is not used for model training.20 Zero data retention by default — Novita's Terms of Service (§10.2) commit that Content is not logged for human review and is not retained beyond the time it takes to generate and deliver the output, except as required by law, as necessary to provide the Services or technical support, or for automated safety screening. No deletion window is stated for the technical-support exception.20 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, and prompts and outputs sent through its model APIs are not used for model training or shared with external model providers, confirmed in writing by Baseten (September 7, 2026; written vendor confirmation).24 Zero data retention by default — prompts and outputs sent through Baseten's model APIs are not retained, used for training, or shared with external model providers; confirmed in writing by Baseten (September 7, 2026; written vendor confirmation, not yet published terms).24 United States (San Francisco HQ). Multi-cloud GPU fleet spans additional regions; serving region for our account: pending verification Subprocessor
GMI Cloud Not published — GMI Cloud publishes no statement on training use of API inputs/outputs (its website privacy policy covers site visitors only). pending verification Not published — no API-content retention or deletion terms are published. pending verification GMI Cloud-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). United States (San Jose, CA company); 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-ultra-nvidia; two further options, nemotron-3-nano-nvidia (retired by NVIDIA effective August 28, 2026) and nemotron-3-super-nvidia, have also been retired by 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. Since September 22, 2026 the Seedance models also accept customer-supplied media inputs (first/last-frame and reference images by URL or inline, reference video and audio by URL), and since September 23, 2026 the Seedream image model accepts customer-supplied input images for edits; all of these are forwarded to ByteDance in transit with the prompt. 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)
StreamLake (Kuaishou "Vanchin" international platform)
Opt-in only — organization PRC and training consents both required		 Yes — locked by default for every organization. StreamLake's international Terms of Service (§4.6) expressly license StreamLake to use Input and Output to train and improve its models, and no account-level opt-out is published. StreamLake-served models (the -streamlake suffixed serving variants) are labeled "China-based" (Kuaishou-parented operator) and serve only organizations whose administrators have enabled both the PRC-hosted-models opt-in and the training consent (see "Organization consent gates" below). Retained "as long as necessary" per StreamLake's privacy policy — no published fixed deletion window and no published retention opt-out. Singapore (stated data-storage location per StreamLake's privacy policy); operator's parent Kuaishou is headquartered in the PRC and cross-border transfers to group entities are permitted; inference hosting location unstated. Subprocessor (only after both organization consents)
Higgsfield Inc. (Wan 3.0 video)
Training and data-retention consents required		 Permitted under standard terms. Terms of Use §4.4 permit using inputs and outputs to train and improve models. Enterprise agreements can provide a no-training carve-out; no such agreement or API-wide exemption is verified for our account. The three -higgsfield video routes require both existing organization consents, not a separate launch flag.29 API outputs remain available at least seven days — this is an availability floor, not a deletion deadline or zero retention. API-specific input/output deletion controls and retention bounds for our account remain pending verification. General account-deletion terms include a 30-day restoration period, backup/legal exceptions, and no removal from already-trained models.29 United States (Higgsfield Inc., San Francisco); serving location and downstream AI/compute providers for our account pending verification. Not classified as a PRC operator merely because it offers Wan models. Independent controller under standard training-permitted terms; listed ahead of deployment and live verification (AIR-1429).
Video Rebirth Limited (BACH by Video Rebirth, bach-1.0 video)
Opt-in only — organization PRC and data-retention consents required		 Not published — Video Rebirth publishes no terms on training use of API inputs or generated outputs pending verification. Its corporate privacy policy and website terms expressly exclude the BACH products and API ("governed by the specific privacy terms associated with those services"; a "separate Product Terms of Service"), and no such product terms are published. Not used for AI Reserve model training in any case (see training).30 Not published — no retention window, deletion control, or deletion-on-request commitment for request inputs (prompts and customer-supplied images by URL) or generated videos pending verification. The API documentation only advises callers to transfer and store generated videos promptly and states that task results can expire; an expiring result link is not a deletion commitment. Behind the data-retention consent gate until terms are published or confirmed in writing.30 Headquartered in Singapore per the operator's own press releases and its website terms (Singapore governing law and courts); its contact page lists Hong Kong / Singapore / United States offices. Legal entity name pending verification — its corporate policies say "Video Rebirth Limited" while bach.art's structured data says "Video Rebirth Private Limited". Processing region for the API pending verification — the api-gen-na endpoint hostname is a hint, not a commitment. Founded by a former Tencent distinguished scientist per its press releases, with Simplified-Chinese API server-error copy and unverified upstream affiliations, so conservatively labeled China-affiliated until verified (the WaveSpeedAI posture): served only to organizations that have not opted out of PRC-affiliated models. Subprocessor only after BOTH organization consents; listed September 29, 2026 ahead of deployment, pricing, and live verification (AIR-1430) — the three bach-1.0-* routes stay unavailable until credentials and pricing are recorded.
Venice.ai (Wan 3.0 video — text-, image-, and reference-to-video — under Venice's own model ids)
Opt-in only — organization PRC and data-retention consents required		 No — Venice's Terms of Service §10.2 (June 8, 2026) state that Venice "shall not use, and shall require that third-party large language model providers … do not use, any User Content to train" any model. Published contractual terms covering the upstream provider as well as Venice; not yet confirmed in writing for our account. Not used for AI Reserve model training in any case (see training).33 Venice itself: "we will not have access to or store the content of your Prompts or Outputs" and temporarily stores generated videos only until download (privacy policy); the platform deletes that copy programmatically after delivery (/video/complete) and re-hosts the clip in its own transient storage behind a 24-hour signed link. Upstream copy: the privacy policy states a "zero data retention policy with our model providers", while Venice's privacy-architecture page says of the anonymized tier these models run on that "the model provider processes your prompt under its own retention and privacy policies" and users should "assume the provider is storing your content" — conflicting published statements, no deletion window, nothing confirmed for our account pending verification. Behind the data-retention consent gate until confirmed in writing.33 Venice's Service is hosted in the United States per its privacy policy (Wyoming governing law per its Terms); legal entity name pending verification. The Wan 3.0 model is Alibaba's closed model with no public weights, so Venice's anonymized tier forwards the prompt and media to a PRC-affiliated upstream for processing; the upstream's location is not published. The platform therefore labels these models China-based and serves them only to organizations that have not opted out of PRC-affiliated models; additionally they are allowlist-restricted, adult-content-gated, third-party-policy-gated, behind the data-retention consent gate, and never publicly listed. Subprocessor only after BOTH organization consents; listed October 14, 2026 ahead of credential and live verification — the three routes answer a configuration error until the API key is recorded.
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
Hugging Face, Inc. (Inference Endpoints) No — Hugging Face states data sent to Inference Endpoints stays within the dedicated deployment and is not used to train or improve models.25 Endpoints serve dedicated GPU deployments that enterprise administrators create one-click from the AI Reserve console; they are separate from gateway chat routing and receive traffic only for the deploying organization. Not stored — Hugging Face states it does not store request payloads or tokens passed to an Inference Endpoint; prompts/completions exist in the dedicated endpoint only while the request is processed. Endpoint access-log and metrics retention for our account: pending verification.25 United States — Hugging Face, Inc. is a US (New York) company; endpoints in the current deploy catalog provision on AWS us-east-1 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 DeepInfra in the United States under DeepInfra's zero-retention terms (they moved from Fireworks AI on September 25, 2026, when Fireworks retired its serverless deployments), even though the underlying models were published by DeepSeek and Moonshot. If DeepInfra degrades, deepseek-v4-flash can fail over to its same-model US backup route on Baseten — the response's x-aireserve-served-model header always names the route that actually served, and that provider's terms above apply to that request. 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-v4.1-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. Novita AI retires its DeepSeek V4 Flash serving on October 9, 2026 (16:00 UTC); deepseek-v4-flash-novita keeps working until then and no fallback chain routes onto it. 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 using Bedrock's global cross-region inference: the request is made from our us-east-1 deployment and AWS may process it in any of its commercial regions. Models with a -bedrock-us suffix route through AWS Bedrock in the United States only (Bedrock's US geographic inference profile). In both cases AWS operates the model inside AWS infrastructure; the model's publisher (Anthropic, Meta, Mistral, DeepSeek) does not receive prompts or completions. Our account's data-retention mode is aws_review (not the legacy provider_data_share mode): where a model's publisher requires it as a condition of access — today Claude Fable 5 and Fable 5.1 — AWS itself may retain and review the content for up to 30 days inside AWS; no mode we have enabled shares content with a model publisher. The Bedrock row above states the per-model effect.

Since August 18, 2026 the Meta Llama and Mistral default names (llama-4-maverick, llama-4-scout, and mistral-large) are additionally served directly on AWS Bedrock alongside their OpenRouter route: the gateway load-balances each request across the two deployments of the identical model weights, and every individual request is processed entirely by one of them — AWS Bedrock's terms (row above) or OpenRouter's terms (row above) apply to that request accordingly. Meta and Mistral never receive these requests on either route. The -bedrock suffixed names remain available to pin a request to the Bedrock route exclusively, and -bedrock-us names to the United States-only Bedrock route.

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. By default that chain contains only deployments of the identical model weights hosted by a second provider on this page — for example, Anthropic Claude and Meta Llama models on AWS Bedrock — so the response still comes from the exact model selected, processed by a different disclosed subprocessor. The platform default never substitutes a different model: cross-model entries (for example, a smaller model in the same family) exist only where a client administrator has explicitly configured them in their own organization's fallback chain. 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. Model names with no second same-weights host carry no default failover chain at all (as do -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 without a verified account-level opt-out (Higgsfield under its standard terms, DeepSeek direct API, Moonshot AI, StreamLake — whose international Terms of Service §4.6 expressly license it to train on Input and Output, covering the -streamlake serving variants — 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, GMI Cloud, Higgsfield's API-account retention controls, Video Rebirth's unpublished BACH API terms, ElevenLabs' default request history for the eleven-* text-to-speech models — Zero Retention Mode is enterprise-only there, and Tako's query retention for the tako-search live-search models — Zero Data Retention is enterprise-only there too). Higgsfield requires both training and retention consents; those consents are not a claim of PRC jurisdiction.
  • PRC-server consent unlocks models served by providers operating under PRC jurisdiction (DeepSeek direct, Moonshot, Alibaba Cloud Model Studio, ByteDance BytePlus ModelArk, StreamLake — Kuaishou-parented operator serving the -streamlake variants) and providers conservatively labeled China-affiliated pending verification (WaveSpeedAI; Video Rebirth's bach-1.0 video routes, which also require the data-retention consent).

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). Each provider's default retention posture is the one stated in its row above — for example, Fireworks AI's row documents zero data retention by default per its published terms, confirmed in writing on September 8, 2026, and Baseten's row carries the zero-data-retention-by-default posture for its model APIs confirmed in writing on September 7, 2026. 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, and audit.

Usage classification (optional, per organization). When enabled, the platform categorizes AI spend by business function for the analytics dashboard by classifying the standing purpose of each API key and team from its name, description, and organizational context (details your organization sets itself, like team names, structure, and which user or team a key belongs to) — one internal model call per key or team when it is created or edited, not per request. By default, classification does not read prompt or response content. An earlier, retired version of this feature classified individual messages; text snapshots it captured stay inside the platform, are never used to train foundation models, and are included in every purge and deletion described below. Classification is off for newly created organizations until an administrator enables it, and administrators can turn it off entirely at any time.

Automatic model routing. When you select Auto and your organization has not turned the Prompt Router off, the platform may send a bounded excerpt of your latest message (up to 4,000 characters, preserving its beginning and end), your organization's routing instructions and quality/cost/speed preferences, model catalog facts, and structural request facts such as tool presence, modalities, output limit, and reasoning effort to TypeSafe AI's Jev model. Jev resolves an explicitly requested model or chooses an eligible model; it does not generate the response to your task. Jev is the only model used for this selection; when it is unavailable or not configured, a deterministic default applies and no other AI provider receives the excerpt. For tool-result continuation turns, the excerpt is the preceding user request, never tool output. Routing excerpts are processed in memory and are not added to routing logs; those logs contain model, usage, timing, and decision metadata. This platform-paid routing decision is separate from the optional, organization-billed business-function classification below. Provider retention policies in the table above still apply. Choosing a concrete model bypasses this Auto selection step.

Per-request classification (optional, separate opt-in). An administrator may additionally turn on per-request classification. Each successful request is then classified from a bounded excerpt of its prompt — the latest user message (at most 2,000 characters), the head of the system prompt (at most 600 characters), and the names of any declared tools — together with the key and team context above, sent to TypeSafe AI's Jev model (see its row in the provider table) through the platform's own LiteLLM engine. The model returns a typed category and probabilities, never generated text. The excerpt is processed in memory and is never stored: not on the classification record, not on the usage record, not in logs; the administrator review queue shows the classification rulebook's question, never the prompt. Each classifier call is metered and billed to the organization as its own usage record (model jev-1.13.0, purpose business-function-classifier), visible in the organization's usage ledger. Per-request classification is off for every organization until an administrator explicitly turns it on, and cannot be on while classification is off.

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 — only the billing metadata above is retained. Usage classification, which reads no message content, is governed by its own toggle above.

Deletion and retention windows. Previously stored content — prompt/response messages and the legacy 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, including its Gemini File Search store (see saved documents below). Billing metadata is retained in each case. See our Terms of Service and Data Processing Addendum for the platform's retention commitments.

Saved documents and Google Gemini File Search. Documents saved to the web document library (from the documents page, or with "Store for future chats" after attaching one in chat) are stored in our Google Cloud project, and the document-processing service sends each one to the Gemini API (paid tier, the row above) to categorize it and extract searchable text. To let chat search the library, the document-processing service also uploads each saved document into a Gemini File Search store scoped to your organization, where Google chunks, embeds, and stores its content together with labels such as your organization, the document, its uploader, the user it was uploaded for, its type, and a fingerprint of its content. Chat questions that search the library are sent to the Gemini API with that store attached and limited to the documents the asking user may read; which documents those are is decided by the platform's own records, not by the labels stored with the copy. The indexed copy has no expiry — Google keeps it until it is deleted27 — and is governed by the Gemini API terms in the row above; the Gemini API offers no region selection, so the row's processing-region statement applies. The platform deletes the indexed copy when the document is deleted (by the user, an administrator, self-service data deletion, or account deletion) and confirms against the store that the copy is gone before the document's record is cleared; a daily sweep removes any indexed copy a deletion could not complete, the organization-level content purge above also removes the indexed copies of already-deleted documents, and an organization's whole store is deleted when the organization is closed. Clean-up of an earlier defect: an earlier defect kept these deletions from completing (the platform had recorded the indexing operation rather than the indexed copy's own identifier). The correction records the copy's own identifier and is being rolled out together with a one-time reconciliation of each organization's store: the daily sweep removes the affected copies whose document records remain, the reconciliation removes the rest, and until it has completed an indexed copy of a deleted document may still be in the store — it is deleted on request. API traffic never adds to the library.

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.

Slack Connect customer channels

Separately from the model providers above, AI Reserve staff can create a Slack Connect shared channel with a customer organization from the AI Reserve admin console. When this is used, Slack Technologies (a Salesforce company) receives the customer admin's email address (used to send the Slack Connect channel invite), a shared channel named for the customer organization, and the messages participants choose to post in that shared channel; Slack also receives our operational alert payloads, which may reference customer organization names. For organizations that use the @aireserve mention agent in their Slack Connect channel, prompts submitted in Slack and the model completions returned there are Slack messages: they are processed and retained by Slack, and each side of the shared channel keeps its own copy under its own workspace retention settings (the customer's and AI Reserve's). Content sent through the API directly does not transit Slack. Slack is listed on the Subprocessors page.

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 (listed ahead of first traffic; both serving production traffic since August 18, 2026); Hugging Face (Inference Endpoints) entry added on August 18, 2026 (GPU one-click deploys); Slack (Slack Connect customer channels) entry added on August 21, 2026 (admin-created shared channels with customer organizations) and broadened the same day for the @aireserve mention agent (prompts and model completions exchanged in Slack Connect channels); MuleRouter training-use, task-deletion, and serving-region entries completed on August 26, 2026 from written confirmations by the Alibaba Cloud International partner team (footnote 26) — written vendor confirmation rather than published terms, marked as such on the row. Baseten training-use and retention entries completed on September 7, 2026 from a written confirmation by Baseten (footnote 24) — written vendor confirmation rather than published terms, marked as such on the row; Baseten's serving region for our account remains pending verification. Fireworks AI training-use and retention entries corroborated on September 8, 2026 by a written confirmation from Fireworks AI (footnote 15) — published terms plus written vendor confirmation, marked as such on the row. DeepInfra, Novita AI, and Together AI training-use and retention entries updated on September 13, 2026 to cite the Zero Data Retention commitments in each provider's current Terms of Service (footnotes 14, 19, and 20) — contractual terms rather than a documentation page; Novita AI's serving region for our account remains pending verification. Google Gemini File Search document-indexing disclosure added on September 16, 2026 (footnote 27) — the saved-document indexing flow and its deletion lifecycle, stated from the platform's own code rather than a provider document; the deletion description was updated on September 17, 2026 to describe the corrected deletion lifecycle and the clean-up of copies the earlier defect left behind. TypeSafe AI (Jev) entry added on September 21, 2026 (footnote 28) for the optional per-request usage-classification mode — training posture from TypeSafe's published model documentation; retention and processing region not published, pending verification. ByteDance (BytePlus ModelArk) entry extended on September 22, 2026: the Seedance video models now also accept customer-supplied media inputs (keyframe and reference images, reference video and audio), forwarded in transit with the prompt under the same PRC consent gate. Higgsfield Inc. added September 22, 2026 (footnote 29), ahead of deployment and live verification: standard training license, unverified API-account retention controls, and United States operator identity are disclosed separately. Video Rebirth Limited (BACH) added September 29, 2026 (footnote 30), ahead of deployment, pricing, and live verification: no published API-content terms, so training use, retention, deletion, and processing region are stated as not published and pending verification; Singapore headquarters is the operator's own statement, not a verification. ElevenLabs added October 11, 2026 (footnote 31), ahead of deployment and account verification: the four eleven-* text-to-speech models dispatch input text to ElevenLabs; the Privacy Policy's training use with its account-level opt-out, the default request-history retention with enterprise-only Zero Retention Mode, and the U.S. default storage location are each stated from ElevenLabs' own published pages. Venice.ai added October 14, 2026 (footnote 33), ahead of credential and live verification: the three Wan 3.0 routes under Venice's own model ids forward prompts and media to a PRC-affiliated upstream on Venice's anonymized tier; the Terms of Service §10.2 no-training commitment, the conflicting published retention statements (privacy-policy zero-retention-with-providers versus the anonymized-tier caveat), and the U.S. hosting statement are each stated from Venice's own published pages. Tako Inc. added October 12, 2026 (footnote 32), ahead of deployment and account verification: the tako-search and tako-search-deep live-search models send only the search query to Tako; the Privacy Policy's paid-plan exclusion from training use, the query retention with enterprise-only Zero Data Retention, and the U.S. operator are each stated from Tako's own published pages. Cloudflare added October 14, 2026 (footnote 34), ahead of deployment: the clef and clef-flash decision models send the request state, questions and any embedded images to Cloudflare Workers AI; the no-training commitment, the storage-only-with- a-storage-product statement and the U.S. operator are each stated from Cloudflare's own published Workers AI data-usage page. 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 — Terms of Service (§2, Zero Data Retention): https://www.together.ai/terms-of-service; 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. Zero-data-retention-by-default posture confirmed in writing by Fireworks AI (September 8, 2026) for its inference API serving open models: prompts and outputs exist only in volatile memory for the duration of the request (with prompt caching, for a few minutes longer) and are never persisted; no training on customer data unless the customer explicitly opts in to a feature such as FireOptimizer; no sharing with external model providers (inference runs entirely on Fireworks infrastructure); only operational metadata (token counts, latency, errors) is logged, never content. Written vendor confirmation corroborating the published terms above. Scope note from those terms: Fireworks' separate Response API endpoint retains conversation data for 30 days when store=true (its default) — the gateway's Fireworks routes use Fireworks' chat-completions inference API, to which the zero-retention default applies.
  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 — Terms of Service (§7(b), Zero Data Retention; §7(a) precedence over the privacy policy and documentation): https://deepinfra.com/terms; Data privacy docs: https://docs.deepinfra.com/account/data-privacy
  20. Novita AI — Terms of Service (§10.2, Zero Data Retention): https://novita.ai/legal/terms-of-service; 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/. Zero-data-retention-by-default posture confirmed in writing by Baseten (September 7, 2026): prompts and outputs sent through its model APIs are not retained, used for model training, or shared with external model providers. Written vendor confirmation, not yet reflected in Baseten's published documentation.
  25. Hugging Face — Inference Endpoints security & compliance: https://huggingface.co/docs/inference-endpoints/security; Privacy Policy: https://huggingface.co/privacy
  26. MuleRouter — confirmed in writing by the Alibaba Cloud International partner team (MuleRouter is an Alibaba Cloud–partnered platform): processing region Singapore (August 25, 2026); no request data retained for training, and AI Reserve holds the right to delete tasks run on MuleRouter (August 26, 2026). Written vendor confirmations, not yet reflected in MuleRouter's published documentation.
  27. TypeSafe AI — Models (Jev 1.13; customer data does not influence model training; customization through request parameters only; pricing and limits): https://docs.typesafe.ai/models; System One concepts: https://docs.typesafe.ai/concepts/system-one
  28. Google — Gemini API File Search (stores hold the chunked, embedded content of uploaded files; no time-to-live — the indexed data persists until deleted): https://ai.google.dev/gemini-api/docs/file-search; File Search Stores API reference: https://ai.google.dev/api/file-search/file-search-stores
  29. Higgsfield — Terms of Use (§4.4 training license and enterprise carve-out; §11 developer access; §16.5 account deletion); Privacy Policy (Higgsfield Inc., US; service improvement and international transfers); API billing and retention (outputs available at least seven days, not deleted after seven days). Reviewed September 22, 2026; no enterprise no-training agreement, API-specific deletion control, processing region, or certification verified for our account.
  30. Video Rebirth (BACH) — bach.art API documentation (General guide: Bearer-token authentication, asynchronous task flow, "transfer and store the generated video promptly", task results can expire; no retention, deletion, training-use, or processing-region terms); corporate Privacy Policy and Website Terms of Use (both dated March 23, 2026, Video Rebirth Limited; expressly limited to the corporate website — the BACH products and API are said to be governed by separate product terms that are not published; website terms choose Singapore law and courts); contact page (Hong Kong / Singapore / United States offices); the operator's press releases (Singapore headquarters; founder formerly a Tencent distinguished scientist). Reviewed September 29, 2026; no API-content terms, processing region, legal entity name, or certification verified for our account.
  31. ElevenLabs — Privacy Policy ("Research and develop our AI models: We use your personal information to research, build, train and improve our AI models"; opt-out "by navigating to the 'Data use' menu in the 'Terms and Privacy' section of your ElevenLabs account", applying to use after the opt-out); Zero Retention Mode (Enterprise) ("By default, ElevenLabs retains data … to enhance services, troubleshoot issues, and ensure the security of ElevenLabs' systems"; ZRM deletes TTS text input and audio output immediately and "is available to select enterprise customers"); Data residency ("As a standard, ElevenLabs' customer data is hosted/stored in the U.S."; processing may occur outside the storage location by international affiliates and subprocessors); API pricing (USD per 1,000 characters). Reviewed October 11, 2026; the account-level training opt-out is set at account creation and its verification on our account is pending.
  32. Tako Inc. — Privacy Policy (last updated August 25, 2026: "We use Queries to provide the Services and to evaluate and improve them, including by training the search and ranking models that power the Services. This doesn't apply to Queries from customers on a paid plan or whose agreement with us excludes it."; "The LLM providers we send Queries to operate under zero-data-retention agreements, and they don't train on your Queries."; "Zero Data Retention. Customers on an enterprise agreement can arrange for us to not retain their Queries."); Search API reference (one POST /api/v3/search call per search; per-request usage.total_cost_usd metering). Reviewed October 12, 2026; listed ahead of deployment, the account's paid-plan status pending verification.
  33. Venice.ai — Terms of Service (last updated June 8, 2026; §9.1 "You authorize Venice.ai to access and exchange your Inputs with Third-Party Services on your behalf"; §10.2 Venice "shall not use, and shall require that third-party large language model providers … do not use, any User Content to train" any model; §20.2 Wyoming governing law); Privacy Policy (effective June 2, 2026; "we will not have access to or store the content of your Prompts or Outputs. We will, however, temporarily store any videos you generate until you download them"; "We operate a zero data retention policy with our model providers"; "The Service is hosted in the United States"); API privacy modes and Privacy in Venice (the anonymized tier: "Venice hides your identity from the provider, but the provider may still see the prompt"; "Provider policies apply … Venice does not control this layer"); Video models (the three Wan 3.0 ids, privacy tier Anonymized) and Video quote API (per-clip USD pricing). Reviewed October 14, 2026; no written confirmation of upstream retention, legal entity name, processing region, or certification for our account.
  34. Cloudflare, Inc. — Workers AI data usage (last updated April 21, 2026: inputs, outputs, embeddings and training data "constitute Customer Content"; "Cloudflare does not make your Customer Content available to any other Cloudflare customer"; "Cloudflare does not use your Customer Content to (1) train any AI models made available on Workers AI or (2) improve any Cloudflare or third-party services, and would not do so unless we received your explicit consent"; "Your Customer Content for Workers AI may be stored by Cloudflare if you specifically use a storage service (e.g., R2, KV, DO, Vectorize, etc.) in conjunction with Workers AI"; processing subject to Cloudflare's Privacy Policy and Self-Serve Subscription Agreement); Clef model page (Apache-2.0 open-weight decision model; per-input-token pricing, output tokens not billed; media tokenized into input tokens); Clairic model page (typed extraction model in research preview on Workers AI since October 14, 2026, same inference API, data-usage terms and per-input-token pricing; images tokenized into input tokens, capped at 4,096 per image). Reviewed October 14, 2026, Clairic added October 20, 2026; listed ahead of deployment on the platform's self-serve account, no processing-region pin or certification verified for our account.