Skip to content
Sign inBook a call

Providers

Where your prompts go

Every answer an AI gives you is computed on someone's hardware, under someone's policy. We read eleven providers' own privacy policies, terms and API documentation against five requirements, and quoted what they say. Checked 29 September 2026.

The five requirements

These are the requirements we are asking AI providers to meet. Each is scored for an ordinary new self-serve account, because that is what most people sign up for.

  1. 1
    No retention of prompts or outputs
    What you send and what comes back is not kept after the request is answered: not for abuse monitoring, not for review.
  2. 2
    No training on prompts or outputs
    Nothing you send or receive is used to train or improve a model.
  3. 3
    Prompt caching
    Repeated context is cached so it is not paid for in full on every request. Chat-completions APIs keep no state between requests, so each one resends the conversation so far; caching is what keeps that affordable.
  4. 4
    Published cached-token prices, cache hits reported
    The price of a cached token is published, and each response says how many tokens came from the cache, so the bill can be checked.
  5. 5
    Inference on GPUs in the United States
    The request can be kept on hardware in the United States.

The matrix

  • Yes, by defaultTrue for a new self-serve account without changing anything.
  • Opt-inA new account can turn it on itself, sometimes per request or at a higher price.
  • Through salesOnly with approval from the provider's sales team or an enterprise contract.
  • NoNot offered.
  • Not statedThe provider's own pages do not say.
Provider
1
No retention
2
No training
3
Caching
4
Cache pricing
5
US GPUs
By default
Anthropic
2 / 5
OpenAI
3 / 5
Google (Gemini API)
3 / 5
xAI
3 / 5
Moonshot AI (Kimi)
2 / 5
DeepInfra
5 / 5
Fireworks AI
4 / 5
Groq
3 / 5
Together AI
3 / 5
OpenRouter
routes to other providers
2 / 5
Hugging Face (Inference Providers)
routes to other providersnot offered on Privatae
0 / 5

Select any cell to see the provider’s own words behind it.

Your turn

Which providers would you choose?

Loading the poll…

Why voting needs an account

An anonymous poll can be filled by bots, and then it measures nothing. So each account gets one ballot. We identify accounts by email address.

The network will be run the same way. Members will discuss it on a members-only community board, and feature proposals will come from members; we build what gathers support. We will put policy changes to a member poll before they take effect, and we will honor member petitions.

How the community works

Provider by provider

Every verdict with the provider’s own words under it. Where a company’s pages disagree with each other, both are quoted.

Anthropic

2 of 5 by default

United States · California law (Ireland for EEA, UK and Swiss customers)

Privacy policy ↗Terms ↗Data processing agreement ↗Subprocessors ↗
1No retention of prompts or outputs
Through sales

By default API inputs and outputs are deleted within 30 days (longer if flagged for Usage Policy violations); zero data retention exists but must be requested from the sales team and enabled per organization.

5 sources for Anthropic: No retention of prompts or outputs
2No training on prompts or outputs
Yes, by default

Anthropic does not train on API inputs or outputs by default; the exception is data the customer explicitly opts in to share, such as feedback.

3 sources for Anthropic: No training on prompts or outputs
3Prompt caching
Opt-in

Prompt caching is opt-in per request via a cache_control field, lives in memory for a 5-minute (default) or 1-hour TTL, is isolated per workspace, and is listed as ZDR-eligible.

7 sources for Anthropic: Prompt caching
4Published cached-token prices, cache hits reported
Yes, by default

The pricing page lists a cache-hit price per model (0.1x base input on most models), and the API usage object reports cache_read_input_tokens and cache_creation_input_tokens.

3 sources for Anthropic: Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Opt-in

Default inference geo is global; any self-serve request can pin inference to US-only infrastructure with inference_geo: "us" at 1.1x the standard price, on Claude 4.6 and later models only.

5 sources for Anthropic: Inference on GPUs in the United States
Research notes: what we couldn’t verify, and where pages disagree
  • CONFLICT: the Privacy Center article (updated July 1, 2026) says API inputs/outputs are deleted 'within 30 days of receipt or generation', while the API docs retention page says 'Conversation content (your prompts and Claude's outputs) is not retained by default; the exception is Covered Models, which require 30-day retention.' R1 verdict follows the stricter reading (30-day default) and ZDR being sales-only.
  • Enterprise/sales unlocks: Zero Data Retention ('contact the Anthropic sales team'; enabled per organization by the account team), HIPAA readiness, custom terms/volume pricing. ZDR excludes Console/playground, Managed Agents, code execution, and the Covered Models (Fable/Mythos 5.x) unless expressly authorized.
  • Human review: no statement found on the fetched Anthropic pages about whether staff may review API content; recorded as not stated. The retention article only says flagged content is retained up to 2 years (classification scores up to 7 years).
  • Feedback data: the retention article says feedback submissions are retained for 5 years; the training article says opted-in feedback may be used for training.
  • Prompt caching is NOT on unless the request carries cache_control (top-level for automatic breakpoints, or per block). 1-hour cache writes cost 2x base input; 5-minute writes 1.25x.
  • US workspace geo ('us' is the only workspace geo) governs data at rest; inference_geo governs where inference runs. inference_geo is not available through the OpenAI SDK compatibility endpoint.
  • could not fetch: https://trust.anthropic.com/subprocessors (JavaScript-only page; subprocessor names not retrievable).
  • Commercial Terms effective June 17, 2025 (as shown on page); privacy policy effective September 10, 2026; training article dated August 18, 2026.
  • Quote formatting: markdown syntax (backticks, italics, link brackets, table pipes kept only where noted) was stripped from quotes taken from docs pages; wording is unchanged.

OpenAI

3 of 5 by default

United States · California law (Ireland for EEA, UK and Swiss customers)

Privacy policy ↗Terms ↗Data processing agreement ↗Subprocessors ↗
1No retention of prompts or outputs
Through sales

By default, abuse-monitoring logs that may contain prompts and responses are kept for up to 30 days; Zero Data Retention and Modified Abuse Monitoring require prior approval by OpenAI via sales.

6 sources for OpenAI: No retention of prompts or outputs
2No training on prompts or outputs
Yes, by default

API data is not used for training unless the customer explicitly opts in.

2 sources for OpenAI: No training on prompts or outputs
3Prompt caching
Yes, by default

Caching is automatic; organizations without ZDR default to extended 24h retention, which may keep encrypted KV tensors in GPU-local storage for up to 24 hours, while ZDR organizations default to in-memory caching.

6 sources for OpenAI: Prompt caching
4Published cached-token prices, cache hits reported
Yes, by default

The API pricing page has a cached-input column for every model, and usage reports cached tokens in prompt_tokens_details.cached_tokens (Chat Completions) and input_tokens_details.cached_tokens (Responses).

5 sources for OpenAI: Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Through sales

US regional processing (us.api.openai.com) exists, but data residency controls require checking eligibility with sales and cost a 10% uplift on models released on or after March 5, 2026; the default Global geography has no stated processing constraint.

5 sources for OpenAI: Inference on GPUs in the United States
Research notes: what we couldn’t verify, and where pages disagree
  • Pricing was taken from https://developers.openai.com/api/docs/pricing (openai.com/api/pricing returned 403). The quoted gpt-6-sol row columns are: short-context input | cached input | cache writes | output | long-context input | cached input | cache writes | output (per 1M tokens).
  • platform.openai.com/docs/guides/your-data now 301-redirects to developers.openai.com/api/docs/guides/your-data.
  • Enterprise/sales unlocks: Zero Data Retention and Modified Abuse Monitoring (prior approval), data residency controls (sales eligibility check; non-US regions additionally require MAM/ZDR approval and a Modified Retention amendment; UAE needs additional approval), Private Retention with Private Safety Processing, BAA/HIPAA.
  • US region row of the data-residency table: domain prefix us.api.openai.com, regional storage Yes, regional processing Yes, 'MAM or ZDR required' No. Default (Global) processing location is not stated; the only text is a code comment '# No processing constraint.' in the example.
  • Human review: the your-data guide says OpenAI 'may retain and human review customer content' in the Safety Retention case for ZDR/MAM customers, and that CSAM-flagged images are 'retained for manual review' even under ZDR. No explicit statement found on human review of default 30-day abuse logs.
  • Caching on GPT-5.6 and later uses prompt_cache_options.ttl with value '30m' ('At least 30 minutes after the latest write or reuse'), and the your-data guide says this TTL 'controls the minimum cache lifetime, not this maximum application-state retention period' (24h GPU-local storage). For gpt-5.5 and gpt-5.5-pro, setting prompt_cache_retention to in_memory returns an error.
  • Chat Completions and Responses endpoints are marked 'Yes, see below for limitations' for ZDR eligibility in the retention table.
  • Quote formatting: markdown syntax (backticks, italics, link brackets, table pipes kept only where noted) was stripped from quotes taken from docs pages; wording is unchanged.

Google (Gemini API)

3 of 5 by default

United States · California law

Privacy policy ↗Terms ↗Data processing agreement ↗Subprocessors ↗
1No retention of prompts or outputs
No

SCORED = Gemini API paid tier: prompts and responses are logged for 55 days for abuse monitoring, flagged content may be reviewed by authorized Google employees, and Google's own ZDR page says guaranteed zero retention requires switching to Vertex AI; on Vertex AI abuse logging (up to 90 days, only when classifiers flag activity) can be turned off only by an approved exception request form (i.e. 'sales'-type approval).

5 sources for Google (Gemini API): No retention of prompts or outputs
2No training on prompts or outputs
Yes, by default

SCORED = Gemini API paid tier (a Cloud project with an active billing account): prompts and responses are not used to improve Google products; the FREE/unpaid tier (AI Studio and unpaid Gemini API quota) IS used to improve and develop Google products and ML, with human reviewers able to read it, except for users in the EEA, Switzerland and UK where paid-tier data terms apply to free use too. Vertex AI: no training without prior permission.

5 sources for Google (Gemini API): No training on prompts or outputs

  • “When you use Paid Services, including, for example, the paid quota of the Gemini API, Google doesn't use your prompts (including associated system instructions, cached content, and files such as images, videos, or documents) or responses to improve our products, and will process your prompts and responses in accordance with the Data Processing Addendum for Products Where Google is a Data Processor.”

    https://ai.google.dev/gemini-api/terms

  • “Your access to Gemini API is a "Paid Service" only when accessing the API through a Cloud Project associated with an active billing account.”

    https://ai.google.dev/gemini-api/terms

  • “When you use Unpaid Services, including, for example, Google AI Studio and the unpaid quota on Gemini API, Google uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services and machine learning technologies, including Google's enterprise features, products, and services, consistent with our Privacy Policy.”

    https://ai.google.dev/gemini-api/terms

  • “To help with quality and improve our products, human reviewers may read, annotate, and process your API input and output.”

    https://ai.google.dev/gemini-api/terms

  • “Google won't use your data to train or fine-tune any AI/ML models without your prior permission or instruction.”

    https://docs.cloud.google.com/gemini-enterprise-agent-platform/resources/zero-data-retention
3Prompt caching
Yes, by default

Implicit caching is on by default for Gemini 2.5+ models, held strictly in RAM (not at rest), project-isolated, 24-hour TTL, and Google states it does not violate zero data retention; explicit caching (opt-in, v1beta) stores content at rest for a user-set TTL (default 1 hour) and is stated to break a zero-data footprint. Vertex AI: same in-memory 24h cache, which can be disabled per project.

5 sources for Google (Gemini API): Prompt caching
4Published cached-token prices, cache hits reported
Yes, by default

The Gemini API pricing page lists a 'Context caching price' per model (e.g. Gemini 3.8 Flash paid tier: $0.075/1M cached tokens vs $0.75 input through Dec 31 2026, plus $0.50/1M tokens per hour storage for explicit caches), and cache hits are reported in usageMetadata.cachedContentTokenCount (Interactions API: usage.total_cached_tokens).

4 sources for Google (Gemini API): Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Not stated

SCORED = Gemini API: the fetched Gemini API pages offer no way to pin inference to the US and the terms say paid-tier data may be stored or cached in any country where Google has facilities; Vertex AI (self-serve Google Cloud) does offer a US multi-region endpoint (aiplatform.us.rep.googleapis.com) and US regional endpoints, priced higher than the global endpoint for Gemini 3+ models (e.g. Gemini 3.8 Flash input $0.825 vs $0.75 global), while its default global endpoint gives no residency guarantee. Google LLC is a US company; API Terms governed by California law.

6 sources for Google (Gemini API): Inference on GPUs in the United States
Research notes: what we couldn’t verify, and where pages disagree
  • SCORING: all five verdicts score the Gemini API self-serve PAID tier (Cloud project with active billing). Vertex AI (now branded 'Gemini Enterprise Agent Platform' on Google's docs; URLs under docs.cloud.google.com/gemini-enterprise-agent-platform) differs and would score: R1 sales (abuse-logging opt-out by approved exception form; Google Cloud Master Agreement customers exempt by default), R2 meets, R3 meets (in-memory cache can be disabled per project), R4 meets, R5 available (US multi-region/regional endpoints, non-global price higher).
  • Free/unpaid tier (AI Studio and unpaid Gemini API quota): content used to provide, improve and develop Google products and ML; human reviewers may read it (Google says it disconnects it from account/API key/project first). Pricing page row 'Used to improve our products': Free Tier 'Yes', Paid Tier 'No'. In the EEA, Switzerland and UK the paid-tier data terms apply to all Services including free.
  • Gemini API ZDR page lists other retention paths for paid users: Grounding with Google Search and with Google Maps store prompts/context/output for 30 days with no way to disable; Interactions API stores state by default (set store=false); Live API session resumption retains up to 24 hours; File API files stay until deleted or expired; explicit context caches persist for their TTL.
  • Vertex abuse monitoring: prompts are logged only if automated classifiers flag suspicious activity (up to 90 days, same region as project); models designated 'Advanced AI' log all prompts and responses for up to 30 days and opt-out 'may not be possible'. Vertex request-response logging to BigQuery is off by default.
  • Gemini API R1 duration: the Additional Terms say only 'for a limited period of time'; the Gemini API usage-policies page gives fifty-five (55) days. Not contradictory, recorded for completeness.
  • Implicit caching: Google states 'no cost saving guarantee' for implicit caching; explicit caching (beta, v1beta) carries the guaranteed discount plus hourly storage cost.
  • Vertex non-global pricing example (Vertex pricing page, Gemini 3.8 Flash through Dec 31 2026): Global input $0.75 / cached $0.075; Non-global input $0.825 / cached $0.0825 per 1M tokens.
  • Not fetched/verified: cloud.google.com/terms/google-entity (which Google entity contracts for Paid Services), Google Cloud Terms Section 15 governing law, and the Google Privacy Policy body (URL resolves 200; not needed for any verdict).
  • Enterprise/sales unlocks: Google Cloud Master Agreement customers are exempt from Vertex abuse prompt logging by default; Provisioned Throughput, dedicated support and volume discounts via Contact Sales (Enterprise tier on pricing page).

xAI

3 of 5 by default

United States · Texas law (its legal pages now name SpaceXAI LLC)

Privacy policy ↗Terms ↗Data processing agreement ↗Subprocessors ↗
1No retention of prompts or outputs
Opt-in

By default all API requests and responses are stored (encrypted) for 30 days for abuse auditing and authorized personnel may review data for misuse investigations; a team admin can switch on Zero Data Retention self-serve in the Console (team-wide, 'where available'), after which content is never persisted to disk and the 30-day retention does not apply, at the cost of stateful Responses, Files, Collections, Batch and deferred completions.

5 sources for xAI: No retention of prompts or outputs

  • “By default, all API requests and responses are stored on our servers (encrypted at rest) for 30 days for auditing purposes in the event of suspected abuse or misuse. SpaceXAI does not train on this data, and it is automatically deleted after 30 days.”

    https://docs.x.ai/developers/faq/security

  • “No storage of prompts: ZDR ensures that API request inputs (i.e., your prompt) and outputs (i.e., the tokens generated by the LLM) are never persisted to disk. The 30-day audit retention described above does not apply to ZDR-enabled teams.”

    https://docs.x.ai/developers/faq/security

  • “Self-serve: Where available, a team admin can turn it on or off directly from the xAI Console. When enabled, it applies to every API key on the team.”

    https://docs.x.ai/developers/faq/security

  • “If you do not see the ZDR option, you may not be a team admin, self-serve ZDR may not be enabled for your environment yet, or your team may be on a negotiated Enterprise Customer Agreement that governs retention separately—contact [email protected] in that case.”

    https://docs.x.ai/developers/faq/security

  • “A limited number of authorized SpaceXAI personnel may review your data when legally required, such as investigating security incidents and potential misuse of our services.”

    https://x.ai/legal/faq-enterprise
2No training on prompts or outputs
Yes, by default

xAI does not train on API inputs or outputs without explicit permission; the only stated opt-in is that xAI 'may offer free credits' in exchange for permission to train on business data. No separate free API tier with different data terms was found on the fetched pages.

3 sources for xAI: No training on prompts or outputs

  • “SpaceXAI never trains on your API inputs or outputs without your explicit permission.”

    https://docs.x.ai/developers/faq/security

  • “SpaceXAI will not use any User Content to train any foundation models, large language models, or other artificial intelligence systems or to develop any new products, services, or features, subject to disclosures to Customer and Customer-controlled user settings.”

    https://x.ai/legal/terms-of-service-enterprise

  • “No, we do not use your business data, including inputs (prompts) or outputs (answers), for training our models. In some instances we may offer free credits in exchange for you agreeing to permit us to train on your business data.”

    https://x.ai/legal/faq-enterprise
3Prompt caching
Yes, by default

Prompt caching is automatic on all grok language models (an x-grok-conv-id header or prompt_cache_key improves hit rates); the docs say entries can be evicted at any time due to memory pressure, server load or restarts and give no fixed TTL; where the cache physically lives beyond that and whether caching is compatible with ZDR are not stated.

4 sources for xAI: Prompt caching
4Published cached-token prices, cache hits reported
Yes, by default

The pricing page has a 'Cached input / 1M tokens' column (e.g. grok-4.7 $0.50 cached vs $2.00 input under 200k tokens), and cached tokens are reported in usage.prompt_tokens_details.cached_tokens (Chat Completions) and usage.input_tokens_details.cached_tokens (Responses API).

4 sources for xAI: Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Opt-in

The default global endpoint (api.x.ai) does not guarantee a processing region; any team can use the US endpoint (us.api.x.ai/v1) which guarantees US request handling, inference, moderation and retained-data storage, at 10% higher token prices and currently only for grok-4.7 and grok-4.6 (no image, video or voice; Files, Collections and server-side tools are outside the guarantee). SpaceXAI is US-based; terms governed by Texas law.

5 sources for xAI: Inference on GPUs in the United States

  • “https://api.x.ai is the global endpoint. SpaceXAI may route requests between regions for capacity or reliability, so the processing location is not guaranteed.”

    https://docs.x.ai/developers/advanced-api-usage/regions

  • “If you need API request handling and model inference to happen in the United States, use the US regional endpoint, https://us.api.x.ai/v1. It is available to every team, and your existing API keys work on both endpoints.”

    https://docs.x.ai/developers/advanced-api-usage/regions

  • “The US endpoint currently serves two models, grok-4.7 and grok-4.6, and none of the image generation, video generation, or voice APIs. Token usage costs 10% more than on the global endpoint.”

    https://docs.x.ai/developers/advanced-api-usage/regions

  • “SpaceXAI LLC ("SpaceXAI") is a US-based company working on building artificial intelligence tools to accelerate human scientific discovery.”

    https://x.ai/legal/faq-enterprise

  • “This Agreement and any dispute or claim arising out of or relating to it or the Services (including non-contractual disputes or claims) shall be governed by, and construed in accordance with, the laws of the State of Texas without regard to conflict-of-laws principles.”

    https://x.ai/legal/terms-of-service-enterprise
Research notes: what we couldn’t verify, and where pages disagree
  • Branding: every fetched xAI page now says 'SpaceXAI' / 'SpaceXAI LLC' (docs.x.ai and x.ai/legal). The privacy policy's contact line still names 'x.AI LLC'.
  • Scope: the Enterprise Terms state '*Note: These terms are for enterprise (business) users of the SpaceXAI API and related SpaceXAI Services (including Grok) who are at least 18 years old.' The consumer Privacy Policy states it 'does not apply to data that we process on behalf of customers of our business offerings, such as the SpaceXAI API'; the DPA (x.ai/legal/data-processing-addendum) is incorporated automatically when personal data is submitted.
  • CONFLICT (retention wording): Security FAQ says requests/responses are 'automatically deleted after 30 days' (https://docs.x.ai/developers/faq/security), but the Enterprise FAQ says 'Inputs and outputs are automatically deleted within 30 days, unless (a) otherwise agreed in writing, (b) SpaceXAI is legally required to retain them because, for example, they are flagged as potentially violating our Terms of Service or AUP.' (https://x.ai/legal/faq-enterprise, last updated February 25, 2025), and the Enterprise Terms allow retention beyond 30 days when 'reasonably necessary for safety, security, compliance, moderation, abuse prevention, or investigation of suspected violations of this Agreement'.
  • Under ZDR the Enterprise Terms say User Content is deleted 'upon the earlier of (a) one hour after completion of the applicable inference request; or (b) delivery of the response' with no logs or backups retained. Features disabled under ZDR: per-key request logging, stateful Responses API, Files, Collections, Batch API, deferred completions, stored image/video outputs, voice agent history. Every response carries an x-zero-data-retention header ('true'/'false').
  • Other default retention: the stateful Responses API stores responses for 30 days by default ('The responses will be stored for 30 days, after which they will be removed.'); store_messages=False disables it.
  • Enterprise Terms: 'Except when Customer elects to use SpaceXAI's Zero Data Retention-enabled APIs ... SpaceXAI may create and use, for any lawful purpose, de-identified and/or aggregated data derived from Customer's use of the Services'. The Enterprise Terms also require customers not to submit Personal Data except through the ZDR-Enabled API.
  • Prompt caching and ZDR: the ZDR 'features not supported' table does not list prompt caching, but no page explicitly states caching is ZDR-compatible, so R3 ZDR compatibility is 'not stated'. Cache storage medium beyond 'memory pressure' eviction is not stated.
  • US endpoint pricing (pricing page): 'billed at 1.1x the global token rates, a 10% premium'; grok-4.7 US: $2.20 / $0.55 / $6.60 per 1M (input / cached input / output) below 200k prompt tokens. ZDR applies on the US endpoint (team-level). No EU or other regional text endpoint was found on the fetched docs (the speech-to-speech model page lists cluster us-east-1 only).
  • Subprocessor list (x.ai/legal/subprocessor-list): all listed locations US except 'SpaceXAI Subsidiary — Services and support — UK'. DPA: 'SpaceXAI may transfer and process Personal Data on a global basis outside Europe as necessary to provide the Services, including in the United States and other countries where SpaceXAI and its Subprocessors maintain data processing operations.'
  • Enterprise/sales unlocks: negotiated Enterprise Customer Agreements may govern retention separately; contractual data-residency requirements -> [email protected] ('A regional endpoint is not, by itself, a comprehensive data-residency guarantee.'); HIPAA BAA via questionnaire (and requires ZDR-Enabled API). SOC 2 Type 2 stated; Trust Center details only for customers with a signed NDA.

Moonshot AI (Kimi)

2 of 5 by default

Singapore · Singapore law

Privacy policy ↗Terms ↗Data processing agreement: Not publicSubprocessors: Not public
1No retention of prompts or outputs
Through sales

By default input is retained while the account is active; Zero Data Retention exists but only on request for enterprise customers via sales, and even under ZDR automated content-safety reviews apply to all traffic; whether humans review content is not stated beyond limited employee access to user data.

6 sources for Moonshot AI (Kimi): No retention of prompts or outputs
2No training on prompts or outputs
Through sales

The terms let Moonshot use customer content to develop and improve its services unless otherwise agreed in writing, and restricting training requires contacting Moonshot for an enterprise arrangement; no self-serve opt-out is stated.

4 sources for Moonshot AI (Kimi): No training on prompts or outputs

  • “We may use Content to provide, maintain, develop, support, and improve the Services, comply with applicable law, enforce our terms and policies, and keep the Services safe and secure. Customer who requires restrictions on the use of Customer Content for training or improving Moonshot AI models may contact Moonshot AI to discuss available enterprise arrangements or separate written agreements. Unless otherwise expressly agreed in writing, Customer Content may be used for the foregoing purposes.”

    https://platform.kimi.ai/docs/agreement/modeluse

  • “**User content:** The prompts, audios, images, videos, files and other content that you input and generate while using our products and services. This information helps us optimize our models and understand your needs and preferences, so that we can provide you with more accurate services and support.”

    https://platform.kimi.ai/docs/agreement/userprivacy

  • “This includes training and refining our underlying technology, such as machine learning models and algorithms.”

    https://platform.kimi.ai/docs/agreement/userprivacy

  • “By default, the Kimi Open Platform does not use enterprise customer data to train models.”

    https://platform.kimi.ai/docs/guide/zero-data-retention
3Prompt caching
Yes, by default

Context caching is automatic with a default 5-minute TTL (1-hour optional), isolated per organization, not manually clearable, and stored in blocks; whether the cache is held in memory or on disk, and whether it is compatible with ZDR, are not stated.

4 sources for Moonshot AI (Kimi): Prompt caching
4Published cached-token prices, cache hits reported
Yes, by default

The pricing page publishes cached-input (cache-hit) prices for every listed model and the Chat Completions usage object reports cache reads in usage.prompt_tokens_details.cached_tokens (usage.cache_read_input_tokens on the Messages API).

4 sources for Moonshot AI (Kimi): Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Not stated

No option to run inference in the US is documented; the privacy policy says collected information is stored on servers in Singapore and that personal data may be processed there, but the pages do not state where GPU inference itself runs.

5 sources for Moonshot AI (Kimi): Inference on GPUs in the United States
Research notes: what we couldn’t verify, and where pages disagree
  • Page dates: terms 'Last Updated: July 30th, 2026'; privacy policy 'Last Update : April 30, 2025'.
  • platform.moonshot.ai redirects to platform.kimi.ai (the international platform used here). platform.moonshot.cn redirects to platform.kimi.com, a separate China-region platform; the docs say 'accounts, balances, and keys on platform.kimi.ai are isolated from other regional Kimi platforms' (https://platform.kimi.ai/docs/guide/troubleshooting).
  • China-region platform (not in scope for a US self-serve account, recorded for completeness): operator quote '北京月之暗面科技股份有限公司及关联公司（简称"我们"或者"月之暗面"）提供' (my translation: 'provided by Beijing Moonshot AI Technology Co., Ltd. and affiliates'); governing law quote '本协议之订立、生效、解释、修订、补充、终止、执行与争议解决均适用中华人民共和国大陆地区法律。' (my translation: 'The formation, effect, interpretation, revision, supplementation, termination, enforcement and dispute resolution of this Agreement are governed by the laws of mainland People's Republic of China.'); venue '任何一方均有权向北京月之暗面科技股份有限公司住所地有管辖权的法院起诉' (my translation: 'either party may sue in the competent court where Beijing Moonshot AI Technology Co., Ltd. is domiciled'). Training clause: '您授予我们一项免费的使用权，以在法律允许的范围内将您输入输出之内容及反馈用于模型服务优化。' (my translation: 'you grant us a free right of use to use your input and output content and feedback, to the extent permitted by law, for model service optimization.'). Source: https://platform.kimi.com/docs/agreement/modeluse
  • Enterprise/sales unlocks: ZDR ('available on request to enterprise customers'), and restrictions on training use of Customer Content ('enterprise arrangements or separate written agreements'). Enterprise plans are obtained via https://platform.kimi.ai/contact-sales.
  • CONFLICT (possible): the ZDR page says 'By default, the Kimi Open Platform does not use enterprise customer data to train models.', while the terms say 'Unless otherwise expressly agreed in writing, Customer Content may be used for the foregoing purposes' (provide, maintain, develop, support, and improve the Services) and point customers who want training restrictions to 'enterprise arrangements', and the privacy policy lists training as a use of information. The pages do not define 'enterprise customer', so it is not stated whether a self-serve pay-as-you-go account counts. The r2 'sales' verdict follows the terms.
  • Human review of content: not stated. The privacy policy says access to user data is limited to 'specific employees ... when necessary for business purposes'; the ZDR page mentions 'Automated safeguards, including content safety reviews'.
  • Retention for abusive content: 'Violation-related information is retained until the violation is resolved.' (privacy policy).
  • Cache Write (separately billed TTL writes) is supported only on kimi-k3; K2 models list cache-hit and cache-miss input prices on the pricing page (e.g. kimi-k2.6 $0.16 hit / $0.95 miss per 1M tokens). kimi-k3 prices: input $3.00, cache write $3.00 (5m) / $6.00 (1h), cached input $0.30, output $15.00 per 1M tokens. The r4 pricing quotes are the raw table-row source from the .md version of the pricing page (columns in order: model, unit, cache write 5min, cache write 1h, cached input, input, output, context for K3; model, unit, input cache hit, input cache miss, output, context for K2).
  • No public DPA or subprocessor list found on platform.kimi.ai; the privacy policy only lists categories of service providers ('hosting services, customer service vendors, cloud services, ...').
  • The home page https://platform.kimi.ai/ lists third-party inference partners for Kimi models (e.g. DeepInfra, Fireworks, Together, Baseten); those are separate providers and not covered by this file.
  • All platform.kimi.ai pages were fetched via their .md variants (the HTML pages are JS-rendered).

DeepInfra

5 of 5 by default

United States · California law

Privacy policy ↗Terms ↗Data processing agreement: Not publicSubprocessors ↗
1No retention of prompts or outputs
Yes, by default

The terms commit to Zero Data Retention of all Customer Data by default and the docs say request content is held only in memory and not logged, with named exceptions (image-generation outputs stored briefly, bulk inference stored on disk, Google/Anthropic models follow those companies' policies, support cases with written consent, and records required by law or to respond to abuse); human review is not stated.

8 sources for DeepInfra: No retention of prompts or outputs

  • “Provider will not retain, store, or log any Customer Data submitted to or generated by the Services beyond the period strictly necessary to process and return the applicable request, after which such Customer Data is deleted from Provider's systems in the ordinary course of operations ("Zero Data Retention").”

    https://deepinfra.com/terms

  • “The foregoing does not apply to: (i) Customer Data that Provider retains at Customer's written request, or with Customer's prior written authorization, to diagnose or resolve a support or Service issue, which Provider will delete within thirty (30) days after the issue is resolved; (ii) non-content operational metadata, such as request counts, timestamps, latency, and billing data, that does not include Customer Data submitted to or generated by the Services; and (iii) records Provider is required to retain to comply with applicable laws or to detect, investigate, or respond to fraud, security incidents, or abuse of the Services.”

    https://deepinfra.com/terms

  • “By executing a Service Order that references these Terms, or by accessing or using the Services, Customer agrees to these Terms and the applicable Service Order.”

    https://deepinfra.com/terms

  • “When using DeepInfra inference APIs, we do not store the data you submit to our APIs on disk. We only hold it in memory during the inference process. Once inference is complete, the data is deleted from memory.”

    https://docs.deepinfra.com/account/data-privacy

  • “We do not log the content of your requests. We log only metadata useful for debugging: request ID, cost, sampling parameters.”

    https://docs.deepinfra.com/account/data-privacy

  • “**Exception:** Outputs of image generation models are stored for a short period of time to allow easy access (e.g., for the model demo page).”

    https://docs.deepinfra.com/account/data-privacy

  • “When using our bulk inference APIs (submitting multiple requests in a single call), we may need to store data for a longer period, potentially on disk in encrypted form.”

    https://docs.deepinfra.com/account/data-privacy

  • “When using the Google model, Google logs prompts and responses for a limited period solely to detect violations of their [Prohibited Use Policy](https://policies.google.com/terms/generative-ai/use-policy).”

    https://docs.deepinfra.com/account/data-privacy
2No training on prompts or outputs
Yes, by default

DeepInfra does not train on API data by default, except that when a customer uses a Google or Anthropic model through DeepInfra, that company's own training policy applies; no free-tier difference is stated.

3 sources for DeepInfra: No training on prompts or outputs

  • “Provider will not sell Customer Data, and will not use Customer Data to train, fine-tune, or otherwise improve any model, except as necessary to provide the Services (e.g., fine-tuning a model for Customer's use).”

    https://deepinfra.com/terms

  • “We do not use data you submit to our APIs for training models, except when using Google or Anthropic models, where the receiving company's training policy applies.”

    https://docs.deepinfra.com/account/data-privacy

  • “We will not store, sell, or train using this data unless we have your explicit consent.”

    https://deepinfra.com/privacy
3Prompt caching
Yes, by default

Prompt caching is automatic on supported models (KV-cache prefix reuse, per-model and per-account, expiring after inactivity), with an opt-in paid retention window of 5 minutes or 1 hour; where the cache is stored (memory vs disk) and whether it is compatible with the ZDR commitment are not stated.

5 sources for DeepInfra: Prompt caching
4Published cached-token prices, cache hits reported
Yes, by default

The pricing page lists a separate cached-input price per model and responses report cache hits in usage.prompt_tokens_details.cached_tokens (plus cache_write_tokens for retention writes).

4 sources for DeepInfra: Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Yes, by default

DeepInfra says it runs inference on its own infrastructure in US-based data centers and its privacy policy names the United States as where data is processed; no self-serve region parameter is documented, and the location of Google/Anthropic models reached through DeepInfra is not stated.

4 sources for DeepInfra: Inference on GPUs in the United States

  • “DeepInfra runs on our own cutting-edge inference optimised infrastructure, in secure US-based data centers.”

    https://deepinfra.com/

  • “Your data may be transferred, stored, and processed outside of your jurisdiction, specifically in the United States.”

    https://deepinfra.com/privacy

  • “These Terms of Service (these "Terms") govern access to and use of the services provided by Deep Infra Inc., a Delaware corporation, email: [email protected] ("Provider"), by the customer identified in the applicable Order Form ("Customer").”

    https://deepinfra.com/terms

  • “This Agreement and the transactions contemplated hereby, and all Disputes between the Parties under or relating to this Agreement or the transactions contemplated hereby, whether in contract, tort, or otherwise, shall be governed by, and construed and enforced in accordance with, the laws of the State of California, without regard to its rules of conflict of laws, and, as applicable, U.S. federal laws.”

    https://deepinfra.com/terms
Research notes: what we couldn’t verify, and where pages disagree
  • Page dates: terms 'Last modified: August 17th, 2026'; privacy policy 'Last modified: August 15th 2026'; subprocessor list 'Last updated: September 6, 2024'.
  • Enterprise/sales: terms say an executed DPA 'supplements, and does not narrow' the ZDR commitment; DeepCluster (dedicated B300 clusters, 3-5 year terms) says 'We can work with your preferred region.' (https://deepinfra.com/deepcluster). No other sales-only privacy unlock found.
  • Terms say the ZDR commitment controls over docs: 'Notwithstanding the foregoing, the Zero Data Retention commitment in Section 7(b) controls over the Privacy Policy and any other Provider policy, webpage, or documentation, and an executed DPA supplements, and does not narrow, that commitment, which applies to all Customer Data.' (https://deepinfra.com/terms)
  • CONFLICT: the terms' ZDR 'applies to all Customer Data' and controls over documentation, while the data-privacy doc says 'Outputs of image generation models are stored for a short period of time' and bulk inference data may be stored 'potentially on disk in encrypted form' (https://docs.deepinfra.com/account/data-privacy). The r1 'meets' verdict applies to ordinary (non-bulk, non-image, non-Google/Anthropic) inference.
  • CONFLICT (possible): the data-privacy doc says 'Once inference is complete, the data is deleted from memory', while the prompt-caching docs say DeepInfra reuses 'the KV (key-value) cache from previous requests' and can retain a prefix for '5 minutes or 1 hour'. The pages do not say whether the KV cache counts as retained Customer Data.
  • CONFLICT (possible): the privacy policy says 'We will not store, sell, or train using this data unless we have your explicit consent', while the data-privacy doc says Google/Anthropic models are exceptions where 'the receiving company's training policy applies' and 'Google will store the output as outlined in their Privacy Notice'. The subprocessor list (dated September 6, 2024) names only Stripe, Amazon Web Services and Google Cloud Platform, and does not list Anthropic.
  • Human review of content: not stated on any DeepInfra page fetched.
  • Prompt cache retention (5m/1h) is only listed for 'nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B' and 'moonshotai/Kimi-K2.7-Code'; write premium 1.25x (5m) / 2.0x (1h) input (https://docs.deepinfra.com/chat/prompt-cache-retention). Automatic caching is 'available on supported models — check the model page for details'.
  • The only region statement found in the docs ('Instances run in `us-west-2` by default. Accounts subject to EU data-protection rules are placed in `eu-central-1`.') is on the Agents product page (https://docs.deepinfra.com/agents/introduction), not inference; not used for r5.
  • could not fetch: https://trust.deepinfra.com/resources (HTTP 403); https://trust.deepinfra.com/ returned a JS shell with no content.
  • https://deepinfra.com/docs/data redirects (308) to https://docs.deepinfra.com/account/data-privacy; verbatim quotes were taken from the .md version of that page.
  • The r4 pricing quote is a table cell from https://deepinfra.com/pricing on the DeepSeek-V4-Flash-0731 row (columns: context 1024k, input '$0.06 / $0.015 cached', output '$0.18').

Fireworks AI

4 of 5 by default

United States · Delaware law

Privacy policy ↗Terms ↗Data processing agreement ↗Subprocessors (Schedule 4 of the DPA) ↗
1No retention of prompts or outputs
Yes, by default

Zero data retention is the default for inference on open models (only metadata such as token counts is logged); the Responses API is the exception and stores conversations for 30 days unless store=False. (Verdict applies to chat/completions on open models; Responses API with default store=True does not meet it.)

5 sources for Fireworks AI: No retention of prompts or outputs
2No training on prompts or outputs
Yes, by default

Fireworks does not train on customer content; the privacy policy allows it only with explicit opt-in.

2 sources for Fireworks AI: No training on prompts or outputs
3Prompt caching
Yes, by default

Prompt caching is on by default for all models and deployments, held in volatile memory, with per-account separation on serverless; the ZDR page itself describes caching.

4 sources for Fireworks AI: Prompt caching
4Published cached-token prices, cache hits reported
Yes, by default

The serverless price list shows input / cached input / output per model, and the chat completions usage object carries prompt_tokens_details.cached_tokens (dedicated deployments also return a fireworks-cached-prompt-tokens header).

4 sources for Fireworks AI: Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Not stated

A US-only serverless endpoint (us.api.fireworks.ai) serves seven models at 1.5x base price with published per-token prices, but Fireworks' pages don't say whether a new self-serve account may use it, so this is scored 'not stated' — the same rule applied to Together AI. Default serverless placement is not stated, and account-wide residency enforcement is Enterprise-only.

6 sources for Fireworks AI: Inference on GPUs in the United States
Research notes: what we couldn’t verify, and where pages disagree
  • R5 eligibility UNVERIFIED: the US-only serverless page gives no sentence on who may use it (no self-serve or sales statement). The verdict 'available' rests on the endpoint being documented publicly with per-token '(US)' rows on the public serverless price list; it is not an explicit self-serve statement. Supported US models listed: Kimi K3, DeepSeek V4 Flash (0731), DeepSeek V4.1 Flash, GLM 5.2, GLM 5.2 Fast, GLM 5.3, GLM 5.3 Flash (GLM 5.2 Fast US at standard pricing).
  • R5: where default serverless traffic runs is not stated. Privacy policy: "Specifically, we have servers for the Service in the US." Regions page: "Fireworks runs a global fleet of hardware on which you can deploy your models." and "New accounts receive GPU quota for the GLOBAL multi-region only." Marketing pricing page: "Region-restricted deployments are priced at a 1.5x premium. Contact Sales to get started."
  • R1 scope: the ZDR statement on the docs page is for 'open models'; retention for any non-open models is not stated. Terms: ZDR "does not apply to any feature of the Service that, by design, requires the retention or further processing of data beyond the time needed to generate and deliver Output — including, among other things, the Response API and any training, fine-tuning, or agent features." Responses API retention is 30 days with store=True default; store=False or DELETE opts out.
  • R1: abuse-monitoring log retention: not stated beyond the terms' reservation to use 'safety screening tools'. Human review: terms say Content is not logged 'for human review' (with the legal/support carve-out quoted in R1).
  • CONFLICT (cache lifetime): data-handling page says prompt data/KV caches 'can be stored in volatile memory for several minutes'; prompt-caching page says cached prompts stay 'at least several minutes' and 'up to several hours'. Storage medium beyond 'volatile memory' is not stated.
  • R4: the cached_tokens field quote is from the OpenAPI schema on the chat-completions reference (field path usage.prompt_tokens_details.cached_tokens, title 'Cached Tokens'); the prompt-caching guide itself only documents the response headers.
  • HQ source quotes (Terms PDF): "provided by Fireworks.ai, Inc."; DMCA contact "Fireworks.ai, Inc. ... 900 Concar Drive, Floor 5 San Mateo, CA 94402"; "your personal information may be transferred to, and/or processed in, the United States."
  • Enterprise/sales unlocks: account-wide data residency enforcement (Settings > Governances > Data Residency, 'US' region; other regions via sales), additional GPU placement quota beyond GLOBAL, region-restricted dedicated deployments at 1.5x.
  • could not fetch as HTML: https://fireworks.ai/terms-of-service and https://fireworks.ai/dpa redirect (308) to PDFs on cdn.sanity.io; text was extracted locally from the fetched PDFs with pdftotext. trust.fireworks.ai subprocessor list not read.
  • Privacy policy last updated 8/11/2026.
  • Scoring note (2026-09-29): R5 moved from 'available' to 'not stated' after review — the US-only endpoint's docs state no access restriction but never state self-serve eligibility, and Together's equivalent gap is scored 'not stated'.

Groq

3 of 5 by default

United States · California law

Privacy policy ↗Terms ↗Data processing agreement ↗Subprocessors ↗
1No retention of prompts or outputs
Opt-in

Inference data is not retained by default, but Groq may log inputs and outputs for up to 30 days for reliability troubleshooting or abuse investigation; any customer can turn on Zero Data Retention in Data Controls to stop that.

4 sources for Groq: No retention of prompts or outputs
2No training on prompts or outputs
Yes, by default

The Services Agreement bars Groq from training on inputs or outputs unless the customer grants permission.

The source for Groq: No training on prompts or outputs

  • “For the avoidance of any doubt and to the extent permitted by applicable law, Groq is not permitted to use Inputs or Outputs for training or fine-tuning any AI Model Services or other models, unless explicitly granted permission or instructed by Customer.”

    https://console.groq.com/docs/legal/services-agreement
3Prompt caching
Yes, by default

Caching is automatic on supported models (the page lists gpt-oss-20b, gpt-oss-120b, gpt-oss-safeguard-20b), held only in volatile memory, expiring after 2 hours without use; no persistent storage.

4 sources for Groq: Prompt caching
4Published cached-token prices, cache hits reported
Yes, by default

Per-model docs pages publish a cached input price (e.g. gpt-oss-120b $0.075/M vs $0.15/M input), and chat completions report prompt_tokens_details.cached_tokens (Responses API: input_tokens_details.cached_tokens).

3 sources for Groq: Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Not stated

No documented way for a self-serve user to pin inference to the US; the x-groq-region response header only reports which datacenter served a request, while stored customer data sits in US GCP buckets.

4 sources for Groq: Inference on GPUs in the United States
Research notes: what we couldn’t verify, and where pages disagree
  • R4 pricing: https://groq.com/pricing no longer shows any price table (only marketing content). The models table at https://console.groq.com/docs/models shows input/output only (e.g. 'openai/gpt-oss-120b | 500 | $0.15 input$0.60 output'), and some models show 'ContactSales'. The cached input price appears on per-model pages, e.g. https://console.groq.com/docs/model/openai/gpt-oss-120b (Input $0.15, Cached Input $0.075, Output $0.60 per 1M). The '$0.075' quote is the cached-input figure from that page.
  • R3: caching is documented only for the three gpt-oss models listed on the prompt-caching page. ZDR compatibility is not stated explicitly; the page states cache content is volatile-memory only and never in persistent storage.
  • R1: human review of logged content: not stated. Services Agreement: "Groq does not access, use, store, or retain Inputs or Outputs except as necessary to provide the Cloud Services, in accordance with the Customer's permission or instruction, comply with applicable law, ensure the reliable operation of the Cloud Services, or confirm Customer's compliance with the AUP." Features that need retention (batch 30 days, fine-tuning/LoRA until deletion, stored application state) are listed on /docs/your-data and can be restricted per feature in Data Controls.
  • R5: where default inference runs is not stated; the latency docs show an example x-groq-region value of us-east-1. 'Regional / sovereign endpoints' are mentioned on the LoRA page but no page found explaining how to obtain them or whether they include a US-only option. Region/data-centre locations appear only in newsroom/blog posts, which are excluded as sources.
  • Privacy policy (last updated November 12, 2025) excludes Customer Data processed as a processor: "This Policy does not apply to the information that we process as a 'data processor' on behalf of customers"; it states "Groq is located in the United States, and maintains processing operations in various global jurisdictions."
  • Legal index (https://console.groq.com/docs/legal) lists current policies 'last modified June 22, 2026'; Services Agreement last updated June 22, 2026. Contracting entity for Americas: Groq LLC; EMEA: Groq UK Limited; Saudi Arabia: Groq Limited.
  • Enterprise/sales unlocks: not stated for retention or region; some models on /docs/models show 'ContactSales' for price and rate limits.
  • could not fetch: https://trust.groq.com/ (JS-only, title only), so the subprocessor list was not read.

Together AI

3 of 5 by default

United States · California law

Privacy policy ↗Terms ↗Data processing agreement: Not publicSubprocessors ↗
1No retention of prompts or outputs
Opt-in

By default Together stores prompts and model responses for product improvement (no retention period stated); an org admin can self-serve Zero Data Retention by turning that setting off.

5 sources for Together AI: No retention of prompts or outputs
2No training on prompts or outputs
Yes, by default

Training on customer data is opt-in and off by default.

2 sources for Together AI: No training on prompts or outputs
3Prompt caching
Yes, by default

Prompt caching is automatic on serverless and dedicated inference, entries last about an hour best-effort; the ZDR page says request content may sit in volatile memory and temporary caches but is not persisted to non-volatile storage.

3 sources for Together AI: Prompt caching
4Published cached-token prices, cache hits reported
Yes, by default

The pricing page lists a cached input price per serverless model, and responses report cached tokens in usage.prompt_tokens_details.cached_tokens (some models use a top-level usage.cached_tokens instead).

3 sources for Together AI: Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Not stated

Serverless requests cannot be pinned to a region; region pinning needs a dedicated endpoint, and a Together support article says targeting specific data centers is 'only available on Scale and Enterprise tier plans', but no Together page fetched says how a Scale plan is obtained (self-serve or sales), so self-serve US pinning is not stated.

4 sources for Together AI: Inference on GPUs in the United States
Research notes: what we couldn’t verify, and where pages disagree
  • HQ source quotes (https://www.together.ai/terms-of-service): "Together Computer, Inc., a Delaware corporation"; "Address: 251 Rhode Island Street, Suite 200, San Francisco, CA 94103".
  • R1: default retention DURATION for stored prompts/responses is not stated on the privacy policy, terms, or privacy docs. Abuse-monitoring retention: not stated. Human review of content: not stated.
  • R1: ZDR is org-level (org admins in Organization Settings > Privacy), applies from the moment it is enabled, and does not cover request metadata. Enabling ZDR also disables passthrough models.
  • Passthrough models: 'Allow passthrough models' is on by default and forwards prompts to third-party providers under their own data policies (https://docs.together.ai/docs/privacy-and-security). A new account's traffic to those models is governed by the third party.
  • Terms quote on ZDR effect: "If you choose to enable ZDR, you acknowledge that the Company will not use your data and outputs to train models, improve Services, or for other secondary purposes except as needed to provide the Services to you." (https://www.together.ai/terms-of-service, last updated May 19, 2026).
  • R3: where the cache physically lives (GPU/host memory vs disk) is not stated on the prompt-caching page; the only location statement is the ZDR page's volatile-memory sentence. Explicit caching/ZDR compatibility sentence: not stated.
  • R5: the dedicated-endpoint settings docs (https://docs.together.ai/docs/dedicated-endpoints/settings) describe placement profiles/region lists with no tier restriction stated; the support article restricts targeted data centers to Scale and Enterprise tier plans. Whether the Scale tier is self-serve was not verified. Where default serverless traffic runs: not stated. The pricing page has no data-center statement.
  • Enterprise/sales unlocks: dedicated endpoints on targeted data centers (Scale/Enterprise tiers per support article), EU data centers.
  • could not fetch: https://docs.together.ai/docs/privacy-settings (HTTP 404; replaced by /docs/privacy-and-security). https://trust.together.ai/ and /subprocessors render only a title (JS-only).
  • Privacy policy last updated December 17, 2025; it states data 'may be transferred to, and maintained on, computers located outside of your state, province, country, or other governmental jurisdiction' without naming countries.

OpenRouter

2 of 5 by default

United States · New York law · Routes requests to other companies' hardware, whose own policies then apply.

Privacy policy ↗Terms ↗Data processing agreement ↗Subprocessors ↗
1No retention of prompts or outputs
Opt-in

OpenRouter itself does not store prompts or completions unless you opt in to logging, but it is a router: by default a request can go to a provider that stores data, and a new account has to opt in to ZDR-only routing itself, either for the whole account in privacy settings or per request with provider.zdr=true.

11 sources for OpenRouter: No retention of prompts or outputs
2No training on prompts or outputs
Opt-in

OpenRouter says it does not train on inputs or outputs, but by default a request can go to a provider that may train on it; a self-serve user can block that per request with provider.data_collection="deny" or with an account privacy setting.

7 sources for OpenRouter: No training on prompts or outputs
3Prompt caching
Yes, by default

Caching is passed through to the provider. It is automatic on most providers, and Anthropic and Alibaba need explicit cache_control. The cache lives in memory in the provider's datacenter, and OpenRouter says in-memory caching is allowed under ZDR routing.

4 sources for OpenRouter: Prompt caching
4Published cached-token prices, cache hits reported
Yes, by default

Each model page lists a Cache Read price, and every response's usage object reports cache hits in usage.prompt_tokens_details.cached_tokens, alongside cache_write_tokens and a cache_discount field.

5 sources for OpenRouter: Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Opt-in

On the self-serve Business plan (8% platform fee, against 5.5% on Standard), requests sent to us.openrouter.ai are processed only by US provider endpoints. The docs do not say where default traffic runs, because routing load-balances across providers by price and uptime.

5 sources for OpenRouter: Inference on GPUs in the United States
Research notes: what we couldn’t verify, and where pages disagree
  • ROUTER: requests are forwarded to third-party model providers whose own terms apply. Terms: "The applicable Model Terms govern your, your Authorized Users', and your customers' access to and use of the applicable Model to the extent required by the applicable Model Provider." Terms also say: "OPENROUTER MAKES NO REPRESENTATION OR WARRANTY REGARDING ANY MODEL PROVIDER'S DATA HANDLING, RETENTION, TRAINING, SECURITY, AVAILABILITY, OR INTELLECTUAL PROPERTY PRACTICES." (https://openrouter.ai/terms)
  • Self-serve controls, all verified above: per-request provider.data_collection="deny" and provider.zdr=true; account-level ZDR toggles per model group (Anthropic / OpenAI / Google / SpaceXAI / all other models); guardrail-level enforce_zdr_* fields; account setting to exclude providers that train, with separate settings for paid and free models; opt-in 'OpenRouter Use of Inputs/Outputs' for a 1% discount (off by default); opt-in Private Input & Output Logging (off by default).
  • A per-model-group ZDR toggle for Anthropic, OpenAI or Google 'Removes first-party ... endpoints (Bedrock and Vertex remain available)' and so on. In ZDR mode, those models are served through Bedrock, Azure or Vertex instead (https://openrouter.ai/docs/guides/features/zdr).
  • ZDR does not cover plugins or tools: "ZDR enforcement only applies to provider routing for inference requests. It does not apply to [plugins] and tools you choose to enable, such as [web search]." (https://openrouter.ai/docs/guides/features/zdr)
  • OpenRouter's own exceptions to its no-storage position: the DPA allows retention 'where required by law or as necessary to address malicious use' plus 'automated risk classification' before deletion. Terms 6.3(b) allows temporary storage for 'batch or other large-volume request processing'. Terms 6.5 says prompts are sampled for anonymous categorization ('Unless explicitly opted in to prompt logging, we do not store your Inputs after categorizing them'). Human review of API content: not stated for inference traffic. For Files API uploads, the privacy policy says 'our personnel do not access the content of Uploaded Files except as necessary to investigate a suspected violation of our Terms of Service, respond to legal process, or provide support at your request.'
  • If you opt in to prompt logging, the Terms (6.2) grant OpenRouter a 'worldwide, perpetual, irrevocable' license, including to 'license or sell your User Content in anonymized form'.
  • Scope: the Free plan (no credit card, 50 requests/day) has 'Prompt Caching: No' and 'Data Policy-Based Routing: No' on the pricing table. A credit-card self-serve account is on the 'Standard' plan (5.5% platform fee), where both are 'Yes'.
  • CONFLICT: the pricing FAQ says "Zero Data Retention routing is available on every plan", but the pricing table row 'Data Policy-Based Routing' shows 'No' for the Free plan (https://openrouter.ai/pricing). This does not affect a paid self-serve (Standard) account.
  • CONFLICT: the Provider Logging page says "For enterprise customers, OpenRouter supports in-region routing in the EU and US... This feature is only enabled for enterprise customers by request." (https://openrouter.ai/docs/guides/privacy/provider-logging). The In-Region Routing doc and the pricing page say it is on the Business plan, upgradable from settings with 'no sales conversation' (https://openrouter.ai/docs/guides/features/in-region-routing, https://openrouter.ai/pricing). The provider-routing page repeats the 'for enterprise customers' wording. The R5 verdict uses the newer, more specific Business-plan pages.
  • Enterprise/sales unlocks: invoicing, SSO/SAML, contractual SLAs, 'Fee discounts available', and a higher BYOK allowance ('$200,000/month with no fee, 5% after'). US/EU in-region routing is also on the self-serve Business plan.
  • BYOK: 'If your own provider agreement includes zero data retention, you can declare that on the key.' Declared ZDR is 'not independent verification by OpenRouter' (https://openrouter.ai/docs/guides/features/zdr).
  • Retention and training status per provider is a live table on the provider-logging page, loaded client-side from https://openrouter.ai/api/frontend/v1/all-providers. The current ZDR endpoint list is at https://openrouter.ai/api/v1/endpoints/zdr. Neither was captured here.

Hugging Face (Inference Providers)

0 of 5 by default

United States · New York law · Routes requests to other companies' hardware, whose own policies then apply. · Not offered on Privatae.

Privacy policy ↗Terms ↗Data processing agreement: Enterprise plan onlySubprocessors (Section 10 of the privacy policy) ↗
1No retention of prompts or outputs
Not stated

Hugging Face says it does not store request or response bodies when routing and keeps debug logs for up to 30 days with no user data, but retention at the downstream provider is left to each provider's policy, and Hugging Face describes no ZDR option or retention-based routing filter.

4 sources for Hugging Face (Inference Providers): No retention of prompts or outputs
2No training on prompts or outputs
Not stated

Hugging Face says it does not store user data for training, but training by the downstream provider is left to each provider's policy, and Hugging Face describes no training-based routing filter.

2 sources for Hugging Face (Inference Providers): No training on prompts or outputs
3Prompt caching
Not stated

The Inference Providers docs do not document prompt caching; the only mention of a cache is in the provider-integration billing docs.

The source for Hugging Face (Inference Providers): Prompt caching
4Published cached-token prices, cache hits reported
Not stated

No cached-token price is published: the Hub API's per-provider pricing object lists only input and output prices. The chat-completion API reference lists only prompt_tokens and completion_tokens in usage, with no cached-token field.

The source for Hugging Face (Inference Providers): Published cached-token prices, cache hits reported
5Inference on GPUs in the United States
Not stated

A user can force a specific provider, or choose by speed, price or preference order, but Hugging Face offers no region or US-only routing option and does not say where default traffic runs; the default picks the fastest provider.

2 sources for Hugging Face (Inference Providers): Inference on GPUs in the United States
Research notes: what we couldn’t verify, and where pages disagree
  • ROUTER: requests go through router.huggingface.co to third-party providers such as Cerebras, Groq, Together AI, Fireworks, DeepInfra, Nscale, Scaleway, OVHcloud, Novita, Replicate, fal and others. Hugging Face defers data handling to them: "please refer to the Data Security Policies of each provider" (https://huggingface.co/docs/inference-providers/security). One provider, 'hf-inference', is Hugging Face's own service: "From a user point of view, working with HF Inference is the same as with any other provider." (https://huggingface.co/docs/inference-providers/pricing)
  • Scoring basis: R1 and R2 are 'not stated', not 'meets'. Hugging Face's own hop does not store bodies or train, but a new account's traffic ends at a third-party provider whose retention and training Hugging Face does not describe, and Hugging Face offers no ZDR or no-training filter like OpenRouter's. If the page scores the router hop alone, the quotes support 'no body storage; logs 30 days without user data; no training'.
  • Self-serve controls that exist: a provider selection policy (:fastest default, :cheapest, :preferred, or a specific provider), a preference order in https://hf.co/settings/inference-providers, and a Custom Provider Key (BYOK, billed by the provider). Team & Enterprise admins can "set a spending limit and disable a set of Inference Providers from the organization's settings" (https://huggingface.co/docs/inference-providers/pricing).
  • The pricing page says 'no markup from Hugging Face'. Free users get $0.10/month in credits and PRO users $2.00/month (https://huggingface.co/docs/inference-providers/pricing).
  • Enterprise/sales unlocks: "Hugging Face can also offer Business Associate Addendums or GDPR data processing agreements through an Enterprise Plan." and "If a contract or specific data storage is something you'll need, we recommend taking a look at our Team & Enterprise Support" (https://huggingface.co/docs/hub/security). None of these pages describes an inference-specific retention or region commitment.
  • The subprocessor list in the privacy policy (effective date March 28, 2023) covers Hub infrastructure and vendors (AWS, GCP, MongoDB Atlas, OVHCloud, Stripe, Metronome, etc.). It does not name the Inference Providers partners. https://huggingface.co/subprocessors returned 404.
  • The Hub is SOC2 Type 2 certified: "The Hugging Face Hub, which Inference Providers is a feature of, is SOC2 Type 2 certified." (https://huggingface.co/docs/inference-providers/security)
  • Human review of inference content: not stated. Abuse-monitoring retention: not stated beyond the 30-day debug logs, which contain 'no user data or tokens'.

How this was done

Only each company’s own pages count: its privacy policy, terms, data processing agreement, API documentation and pricing. No blogs, no aggregators. Every verdict carries the quote it rests on and the address it came from. If a page could not be read, the verdict says “not stated” rather than guessing.

Each requirement is scored for a new self-serve account on the provider’s standard API. What an enterprise contract or a sales conversation can add is scored as “through sales”, because most people will never have one.

Policies change. This page describes 29 September 2026; when we check again, the date changes with it. If we got something wrong, write to [email protected] with the page that shows it, and we will correct it here.

Privatae does not sell inference. You bring your own provider account and pay the provider directly.

What Privatae costs · How we handle your data

Where your prompts go — Privatae