Providers

Tracked as a confidential inference provider with attested provider compute and no prompt/output logging claims.

NEAR AI declares zero data retention for its hosted confidential inference. On TrustedRouter this covers only release-pinned direct Private TEE routes, not NEAR AI's Incognito or Attested Third-Party routes. ZDR describes inference content, not account, billing or operational metadata. Cloud Terms section 7.2 makes retention route-specific; the general privacy policy covers Cloud website and account-administration data, not an all-services ZDR guarantee. TrustedRouter connects directly to the model workload and verifies the live TLS key, Intel TDX quote, NVIDIA GPUs, deployment action log, and pinned workload inside the TrustedRouter enclave before sending content. Verification fails closed.

TrustedRouter encrypts each request to an attested Chutes workload and verifies Intel TDX plus NVIDIA GPU attestation inside the TrustedRouter enclave before sending content. Verification fails closed. Chutes also documents no prompt/output storage or training.

Requests and responses are encrypted between TrustedRouter's enclave and release-pinned Privatemode workloads after attestation verification. Verification failures reject the request; there is no plaintext fallback. Privatemode explicitly states that prompts and responses are not stored after a request completes and are not used for training. Transient inference state may remain in a salt-isolated, in-memory prompt cache inside confidential workers until eviction. Operational metadata is retained for up to 90 days; per-key token usage is retained permanently for billing. Cache-routing metadata is visible to the provider edge. ZDR describes prompt/output retention, not metadata.

Baseten states that it does not store synchronous Model API inputs or outputs by default. TrustedRouter uses the synchronous Model API path; Baseten documents separate temporary input storage for async inference.

Tracked as provider-ZDR. Cerebras documents zero-retention inference and ZDR-compliant ephemeral prompt caching that is never persisted.

Marked ZDR — Cohere does not retain prompt/response content for TrustedRouter's configured account and does not train on customer API data. (Not a confidential-compute/TEE provider.)

Confidential AI (Inexorable, Inc.) publishes zero prompt/completion retention and no training on customer data. Operational and billing metadata are retained. Inference is served in the US and El Salvador. Its public attestation covers launch/admission, not verified GPU state. TrustedRouter uses its HTTPS API; these routes are ZDR, not verified end-to-end confidential inference and are excluded from the e2e filter.

Engy states that prompts, outputs, tool arguments, images, and embeddings are not stored or used for training. It retains request metadata such as model, token counts, cost, and latency. Engy's verified-inference sampling is not a user-verifiable TEE attestation, so this route is ZDR but not confidential or E2EE.

TrustedRouter's managed Vertex AI account is covered by contractual Zero Data Retention. This guarantee applies only to TrustedRouter-funded routes. TrustedRouter does not invoke Google Search or Maps grounding or Gemini Live session resumption on these routes. Google AI Studio is classified separately.

Greenference SAS's DPA commits to EU-only inference, no persistent prompt/output storage and no training or secondary use. Shared prefix caches remain in volatile GPU memory until eviction or instance restart; they are not account-isolated and have no fixed maximum lifetime. Content-free operational metadata is retained. Greenference runs its own serving software on rented EU compute; this is contractual ZDR, not verified confidential compute or E2EE. Its current electricity mix is unverified, not a 100% renewable-energy commitment.

Inceptron documents zero retention by default: API prompts and outputs are processed transiently and discarded after processing.

Marked ZDR via TrustedRouter's arrangement — Nebius RETAINS inputs/outputs by default (for speculative decoding); zero retention is an opt-in control, which the deployed Nebius account has enabled. Nebius does not train on customer data.

Contracted Zero Data Retention is active for TrustedRouter's managed OpenAI account, effective July 28, 2026, and live enforcement was verified on July 29, 2026. This guarantee applies only to TrustedRouter-funded routes; customer BYOK credentials use the data controls on the customer's own OpenAI organization or project.

Tracked as ZDR for serverless and dedicated inference. Parasail documents no storage or logging of submitted input on those service paths, retention only while generating and delivering output, and no training on input or output. Batch service is excluded from this claim; TrustedRouter does not route Parasail traffic through batch.

Phala publishes TDX/GPU attestation and signed request receipts, but TrustedRouter does not yet verify the complete receipt chain on every routed request. The route is therefore not classified as provider E2EE and is excluded from trustedrouter/e2e.

Regolo states that inference prompts and outputs are processed in memory and discarded, with no reuse for training. This is a provider policy, not verified confidential compute or provider-side E2EE.

Carbon ScaleDown, Inc. serves compression, summarization, extraction, and classification tasks. Its public DPA commits to volatile-memory processing, zero retention of customer inputs and outputs, no model training on customer data, and US-only processing. Content-free operational logs may be retained for up to 90 days. These are ZDR routes, not attested confidential-compute or E2EE routes. Input tokens are billable; output tokens are free.

Telnyx's published policy states zero data retention for Telnyx-hosted /v2/ai/openai/chat/completions and /v2/ai/openai/embeddings: content is processed in memory and discarded after the response; request metadata is retained. TrustedRouter currently publishes hosted chat routes and adapts Responses to chat completions. Stateful Responses, AI Assistants with conversation history, third-party models and /v2/ai/embeddings are outside this ZDR scope. OpenAI-compatible embeddings may run in any region. DPA 3.1 restricts training of models available to others, with exceptions for written opt-in or data de-identified under 3.5. This is policy-backed ZDR, not confidential compute or verified end-to-end attestation.

Tracked as provider ZDR. Together documents that inference inputs and outputs are not stored by default; temporary prompt caching may be used for performance, and sharing content for training is opt-in.

Marked ZDR — Voyage AI does not retain prompt content for TrustedRouter's configured account and does not train on customer API data. (Not a confidential-compute/TEE provider.)

Zro declares no retention of prompts, outputs or request bodies and no training or secondary use without separate consent, including by its infrastructure providers. Short-lived, account-scoped prompt caches and content-free operational metadata remain. This is provider-declared ZDR, not independently verified confidential compute or E2EE.

0G labels these routes private, TeeML, TEE-attested, and healthy. TrustedRouter requests that mode with X-0G-Provider-Trust-Mode: private and excludes standard and TeeTLS routes. The current live 0G router path does not expose the route quote, audited code measurement, or response authentication needed for TrustedRouter to verify and encrypt each request end to end, so 0G is not in the trustedrouter/e2e pool.

Abliterate says prompts and completions are discarded after each request, but they transit an upstream inference provider whose retention is not identified in its terms. End-to-end ZDR and confidential inference are unverified.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's Aion Labs account. This route is Standard.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's AkashML account. This route is Standard.

No provider-ZDR claim is tracked here. Alibaba Cloud Model Studio model availability and pricing are linked for users who need to review API data handling and regional deployment scope.

Not currently marked ZDR in TrustedRouter. Anthropic may offer contracted or account-specific data-retention terms, but this provider is excluded from trustedrouter/zdr until that posture is reverified.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's Arcee account. This route is Standard.

Atlas Cloud publishes a platform zero-retention policy, while its aggregated model routes can involve downstream providers. TrustedRouter keeps the public route tier conservative pending downstream-path contractual verification.

Credentials are staged for provider discovery, but no canaried route with exact first-party pricing is published yet. No contractual ZDR, confidential-compute, or E2EE claim is tracked.

Black Forest Labs does not publish contractual zero retention, confidential compute, or end-to-end encryption for the hosted FLUX API. This route is Standard.

Boundless advertises no prompt/output retention or training. Its legal terms permit retention needed to provide, secure, improve or maintain the service and do not warrant third-party infrastructure retention practices. These routes remain Standard, not ZDR, confidential compute or E2EE.

Direct BytePlus ModelArk inference. Availability and standard token prices are refreshed from BytePlus's native catalog and pricing documentation. Video availability requires a verified native adapter.

No provider-ZDR claim is tracked here. Cloudflare's Workers AI documentation is linked for model and data-handling review.

No provider-ZDR claim is tracked here. Crusoe's Managed Inference docs and pricing/catalog pages are linked for model and API data-handling review.

Credentials are staged for provider discovery, but no canaried route with exact first-party pricing is published yet. No contractual ZDR, confidential-compute, or E2EE claim is tracked.

Databricks Foundation Model APIs are a Designated Service and pay-per-token workloads are HIPAA compliant. Databricks may temporarily process or store inputs and outputs for abuse prevention, and may process data outside the originating cloud or region. This route is therefore standard privacy, not ZDR, confidential compute, or end-to-end encrypted.

Decart's public terms permit service-related use of submitted content and do not promise zero retention or confidential compute. This route is Standard.

Typed decision models, not chat completions. Decision Models states that prompts and outputs are processed in memory, never stored or used for training. Operational billing metadata is retained. No confidential-compute or E2EE claim is made. This route enforces the EU tier; inference stays within the EU (currently Finland).

Typed decision models, not chat completions. Decision Models states that prompts and outputs are processed in memory, never stored or used for training. Operational billing metadata is retained. No confidential-compute or E2EE claim is made. Global uses spare EU capacity and may use worldwide capacity for opted-in keys; it is not an EU residency guarantee.

Tracked as no-store, not strict ZDR. DeepInfra documents memory-only handling and no training for ordinary inference, but reserves the right to log a small portion of requests for debugging or security. Google- and Anthropic-backed routes also inherit those vendors' terms.

Not ZDR. DeepSeek's published privacy policy says prompts/inputs may be collected and personal data may be used to train or improve machine learning models and algorithms.

No provider-ZDR claim is tracked here. DigitalOcean's Gradient AI model and pricing documentation is linked for data-handling review.

Native text-to-speech with character-based billing. Standard retention applies; enterprise Zero Retention Mode is not enabled on this route. No confidential-compute or provider-side E2EE claim is made.

TrustedRouter supports bounded fal image generation and MiniMax H3 Max video generation with first-party pricing and paid generation canaries. H3 Max video is returned inline through fal's fixed queue host rather than a provider-selected download host. No contractual ZDR, confidential-compute, or E2EE claim is tracked.

Featherless routes are limited to chat models explicitly available on TrustedRouter's current plan with exact prices in Featherless's authenticated catalog. No contractual ZDR, confidential-compute, or E2EE claim is tracked for this account, so these routes are Standard.

No provider-ZDR claim is tracked here. Fireworks publishes security, privacy, and zero-retention documentation; enable a contracted ZDR posture before marking this provider as ZDR.

No provider-ZDR claim is tracked here. Friendli's legal/privacy terms are linked for users who need to review API data handling.

GMI runs isolated/VPC GPU inference, but that is network isolation, NOT an attested TEE — so no confidential-compute, zero-retention, or E2EE claim is marked. Retention/training terms are unverified (the published policy page is JavaScript-only and would not render).

Not currently marked ZDR in TrustedRouter. Google AI Studio and the Gemini Developer API have product- and billing-specific data-use terms, so this route stays outside trustedrouter/zdr.

Credentials are staged for provider discovery, but no canaried route with an unambiguous downstream provider and exact price is published. No contractual ZDR, confidential-compute, or E2EE claim is tracked.

The configured key was rejected by the documented current and legacy inference catalog hosts with HTTP 401. Inference routes remain disabled pending valid credentials, current per-token prices and successful canaries. No ZDR or confidential-compute guarantee is enabled.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's Inception account. This route is Standard.

In its September 28, 2026 marketplace submission, io.net states that inference prompts and completions are processed only in memory, are not stored or logged, and are not used for training. Content-free operational metadata is retained for billing, capacity planning and abuse prevention. This is a provider declaration: the reviewed public privacy policy does not establish inference-specific zero-retention terms, and the applicable DPA has not been reviewed. Routes remain Standard pending that evidence. Confidential compute is offered on request, but this integration has no verified confidential-compute or provider-side E2EE guarantee.

Jina routes are embedding-only and use exact prices from Jina's authenticated model catalog. No contractual ZDR, confidential-compute, or E2EE claim is tracked for this account, so these routes are Standard.

No provider-ZDR claim is tracked here. Kimi/Moonshot policy source is linked for users who need to review API retention and processing terms.

No provider-ZDR, confidential-compute, or E2EE claim is tracked for the Kling video API. TrustedRouter relays and does not durably store prompt or generated video content.

Krea exposes an asynchronous media API, not a shared OpenAI-compatible chat catalog. TrustedRouter supports Krea 2 Medium with exact fixed per-image billing. The route remains dark until its paid generation canary succeeds; Krea requires a funded API balance, separate from its workspace credits. No ZDR or confidential-compute claim is tracked.

No provider-ZDR claim is tracked here. Lightning's general privacy and security documentation is linked for retention review.

No provider-ZDR, confidential-compute, or E2EE claim is tracked for the LTX video API. TrustedRouter relays and does not durably store prompt or generated video content.

Liquid's configured developer credentials target model deployment and bundling rather than a shared token-priced inference API. The provider remains non-routable until a concrete hosted endpoint is configured.

Lyceum's terms state that inference inputs and outputs are used only to generate results, are not stored beyond technical necessity, and are not used for training or analysis. No account-specific ZDR or verified confidential-compute guarantee is tracked; these routes remain Standard. German company jurisdiction is not a guarantee that every model executes in the EU.

No provider-ZDR claim is tracked here. Makora's inference and privacy documentation are linked for users who need to review API data handling.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's Mancer account. This route is Standard.

TrustedRouter sends requests through its attested gateway to Meta Model API directly. Only Standard-tier Muse Spark models are offered; Meta states these prompts and completions are not used for training. No-training is not zero retention. This route is not marked ZDR, confidential-compute, or end-to-end encrypted.

No provider-ZDR, confidential-compute, or E2EE claim is currently tracked for TrustedRouter's Microsoft Foundry account. Azure model availability and pricing are synchronized directly from the account.

No provider-ZDR claim is tracked here. MiniMax's product privacy overview is linked for users who need to review API/open-platform terms.

TrustedRouter's managed Mistral account has model-training use disabled, so new input and output sent through TR-funded routes is opted out of model training. This is not zero data retention: Mistral's published policy says ordinary API input and output may be retained for 30 rolling days unless ZDR is separately activated. Customer BYOK accounts keep their own Mistral privacy settings.

The configured Modal credentials deploy workloads; they do not identify a shared model catalog. Modal remains non-routable until explicit deployment endpoints and prices are configured.

Morph's public paid tier documents up to 30 days of content retention. Zero retention is reserved for enterprise contracts.

Neurometric states that TrustedRouter requests run with upstream trace logging disabled: prompts and completions are not written to observability or object storage, and only aggregate request counts and token totals are retained. This is classified as no-store, not contractual ZDR or confidential compute.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's NextBit account. This route is Standard.

No provider-ZDR claim is tracked here. Novita's privacy policy says personal information is not used for model training; customer-content processing is governed by customer agreements.

Nscale publishes a general nonlogging statement for serverless inference, but TrustedRouter has not verified a contractual zero-retention commitment for this account. These routes remain Standard and make no confidential-compute or E2EE claim.

NVIDIA's hosted NIM API Catalog endpoints are preview services for development and prototyping. TrustedRouter exposes chat-capable models from its configured API Catalog key as Standard routes. NVIDIA requires an NVIDIA AI Enterprise entitlement for production deployments. NVIDIA does not publish hosted preview token rates, so TrustedRouter uses a conservative fallback price and excludes these routes from price indexing.

TrustedRouter sends requests through its attested gateway to OpenRouter, which routes them to a downstream operator. Routes here are Standard privacy and are excluded from ZDR, confidential-compute, and E2EE routing.

Not routable: the configured OVH credential is not an AI Endpoints access token and the live model catalog rejects it. The provider stays listed while a valid inference token is obtained.

Pearl Research states that it retains operational metadata for security and does not train on customer data. No public contractual zero-data-retention terms are linked, so TrustedRouter classifies these routes as Standard and excludes them from ZDR, confidential-compute, and E2EE routing.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's Perceptron account. Its public catalog is priced, but the configured inference credential did not authenticate during the live canary, so routes remain dark.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's Perplexity account. Sonar runs with low search context and its first-party token and fixed successful-request charges settle exactly. This route is Standard.

Credentials are staged, but no shared canaried catalog with exact first-party token pricing is published yet. No contractual ZDR, confidential-compute, or E2EE claim is tracked.

Recraft's developer terms say API inputs and outputs are not used to train its models, but do not promise zero retention or confidential compute. This route is Standard.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's Reka account. This route is Standard.

Relace's public API documentation does not publish contractual zero retention, confidential compute, or end-to-end encryption for hosted open-model requests. This route is Standard.

Riverflow is a variable-credit media workflow API, not a shared token-priced chat endpoint. Its configured credential is rejected by the live MCP endpoint, and no exact per-job billable receipt is exposed. It remains fail-closed until both contracts are available.

No provider-ZDR, confidential-compute, or E2EE claim is tracked for the Runway video API. TrustedRouter relays and does not durably store prompt or generated video content.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's Sail Research account. This route is Standard.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's Sakana account. Version-pinned routes use Sakana's first-party token prices and authenticated availability.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's SambaNova account. This route is Standard.

Scaleway routes are discovered from its authenticated Model as a Service catalog and priced from Scaleway's first-party EUR price feed with a bounded FX reserve. No contractual ZDR, confidential-compute, or E2EE claim is tracked for this account, so these routes are Standard.

No provider-ZDR claim is tracked here. SiliconFlow's privacy policy source is linked for retention and interaction-data terms.

StepFun's public privacy policy does not provide contractual zero retention or confidential-compute guarantees for API requests. TrustedRouter therefore classifies this route as Standard.

No provider-ZDR claim is tracked. StreamLake's public privacy policy is linked for API data-handling review.

Polyphemus sends the conversation and tool definitions to Telluvian for model selection. No ZDR or confidential-compute claim is made.

TokenHub's Singapore API uses global resource scheduling, not a Singapore-only inference guarantee. Only canaried chat routes with matching regional USD prices are enabled. No verified contractual ZDR, confidential-compute, or upstream E2EE claim is tracked.

No provider-ZDR claim is tracked here. Thinking Machines Lab's model and pricing documentation is linked for model capability and API review.

TrustedRouter calls TypeSafe AI's own API for the Jev decision model. TypeSafe states that it does not train models on customer data. It offers zero data retention to enterprise customers on request; that is NOT configured for TrustedRouter's account, so no ZDR, confidential-compute, or E2EE claim is tracked for this route.

No contractual zero-retention, confidential-compute, or E2EE claim is tracked for TrustedRouter's Upstage account. This route is Standard.

Mixed model-specific posture. TrustedRouter cannot independently verify a complete chain from a live Venice endpoint through immutable source and hardware measurements to committed model weights. Venice is therefore not tracked as confidential or E2EE. Exact routes marked Private in Venice's live catalog qualify only as policy-backed ZDR through endpoint-specific records; Anonymized routes do not.

Fallback route for TypeSafe AI's Jev decision model, through Vercel AI Gateway's /v1/evaluate endpoint. A request served here transits two third parties (Vercel, then TypeSafe). No contractual ZDR, confidential-compute, or E2EE claim is tracked for this route.

Not routable: the configured credential is rejected by both Vultr's account API and Serverless Inference API. Vultr issues a distinct inference key per subscription; the provider stays listed until that key is available.

Wafer supports request-scoped ZDR via Wafer-ZDR: required on supported models; model-level support differs, so TrustedRouter keeps provider-level claims conservative.

W&B Serverless Inference runs on CoreWeave infrastructure. W&B does not publish a zero-retention, confidential-compute, or end-to-end-encryption commitment for this service, so these routes use the Standard privacy tier. TrustedRouter does not enable W&B Weave tracing for provider calls.

xAI documents no training on API requests and 30-day default audit retention, with ZDR as an enterprise feature.

No provider-ZDR claim is tracked here. Xiaomi MiMo's open-platform terms are linked for users who need to review API data handling.

No provider-ZDR claim is tracked here. Z.AI/BigModel policy source is linked for users who need to review API retention and processing terms.