Cloud model platform · Verified August 16, 2026

TrustedRouter vs Microsoft Azure AI Foundry

Microsoft Foundry is Azure's first-party AI platform: 1,900+ models including both GPT and Claude frontier lines, a GA cross-vendor model router, and deep Azure governance. TrustedRouter is a cloud-neutral gateway across ~49 providers with no durable prompt logs on realtime inference and a hardware-attested prompt path you can verify.

Compare details

8decision dimensions

8official sources

1base URL to try TR

Choose Microsoft Azure AI Foundry when

Its operating model is the feature.

Choose Microsoft Foundry when your platform is already Azure and its controls bind: Entra ID and Azure Policy governance, FedRAMP High or HIPAA BAA obligations, PTU reservations and committed spend, or the need for both OpenAI GPT and Anthropic Claude frontier models behind one Azure bill with a GA cross-vendor router.

Choose TrustedRouter when

Privacy needs evidence.

Choose TrustedRouter when you want one API across roughly 49 providers rather than one cloud's catalog, realtime inference that keeps no durable prompt or output logs by default rather than by approval-gated exemption, privacy and region routes (zdr, e2e, eu), and a gateway whose running build you can verify by hardware attestation on GCP, AWS, and Azure.

DimensionMicrosoft Azure AI FoundryTrustedRouter

DeploymentAzure resource deployments: Global, Data Zone (US/EU), regional, PTU, and Batch typesHosted control plane, source-available (BUSL-1.1), with an attested API path

API surfaceAzure OpenAI Chat + Responses at /openai/v1/; Claude via Anthropic Messages APIOpenAI Chat Completions and Responses plus Anthropic Messages

Model access1,900+ models incl. GPT and Claude frontier lines (Claude under Marketplace terms)500+ models across ~50 providers: prepaid, BYOK, and direct

RoutingGA trained router across OpenAI, Anthropic, xAI, DeepSeek, Meta; subsets + failoverProvider fallback plus auto, cheap, fast, free, ZDR, E2E, and EU routes with composable privacy and jurisdiction preferences

ObservabilityAzure Monitor metrics, cost analysis, Foundry tracing, evaluations, dashboardsMetadata analytics and opt-in external broadcast

Prompt contentAbuse monitoring on by default; flagged prompts stored for review, no documented capNo durable prompt or output logs on realtime inference; batch is opt-in encrypted retention

VerificationProduct Terms + Azure audits; ContentLogging:false flag verifiable after approvalLive gateway attestation on three clouds, bound to published source and release evidence

BillingAzure meters + $0.14/1M-input router markup; Claude billed via Marketplace CCUsPrepaid at provider price + 5.5% ($0.01/M floor), or BYOK

What Microsoft Foundry actually is

Microsoft Foundry is Microsoft's first-party AI platform on Azure, renamed twice in two years: Azure AI Studio (2023), Azure AI Foundry (November 2024), Microsoft Foundry (Ignite, November 18, 2025). One Azure resource combines a catalog of more than 1,900 models, the Foundry Agent Service, evaluations, tracing, and governance through Entra ID and Azure Policy. Deployments come in Global, Data Zone (US or EU processing boundary), regional, Provisioned (PTU), and Batch types.

Two commercial regimes share the platform. "Models sold by Azure," including the Azure OpenAI line, are hosted by Microsoft, and the data-privacy documentation commits that prompts and outputs are not available to OpenAI or other model providers. Claude models, generally available since June 29, 2026, are Azure Marketplace offerings under a separate regime: Anthropic acts as an independent data processor for prompts and outputs, and usage bills in Claude Consumption Units.

Where Foundry is the right choice

Foundry is the only cloud platform with both OpenAI GPT and Anthropic Claude frontier models behind one procurement surface, and its trained model router reached GA at Ignite 2025, routing across OpenAI, Anthropic, xAI, DeepSeek, and Meta models (27 in the current pool) with subsets and automatic failover. The compliance portfolio is deeper than anything we offer: Azure OpenAI sits in FedRAMP High audit scope, including Azure Government DoD impact levels, Microsoft signs a HIPAA BAA through the Product Terms, and Azure's ISO and SOC audit portfolio is among the largest of any cloud. Cost machinery: PTU reservations, Batch at 50% off, committed spend, one Azure bill. If your platform is already Azure and those constraints bind, Foundry is the sane default.

The logging models, precisely

This is the load-bearing difference. Azure-sold models run abuse monitoring by default: when automated review (which stores nothing) is inconclusive, flagged prompts and completions go to a per-geography store for human review by Microsoft employees. The historical cap was explicit — an archived 2024 legal page said Azure OpenAI "stores all prompts and generated content securely for up to thirty (30) days" — yet Microsoft removed that language, and the last 30-day mention left the docs in a commit on May 18, 2026. As of August 2026, no maximum retention window is documented for that store. The opt-out, modified abuse monitoring, is "available only to customers and partners managed by a Microsoft account team" or eligible programs; self-service customers cannot turn storage off. One genuine credit: after approval the resource JSON exposes "ContentLogging": "false", a machine-readable off switch most providers do not offer.

TrustedRouter's default is the inverse. Ordinary synchronous and streaming inference never logs prompt or output content. We retain metadata: request ids, model and provider, token counts, latency, cost, region, API-key hash. Batch is a separate opt-in mode with enclave-encrypted retention up to 30 days (details). There is no approval process because there is no storage to turn off.

Trust by contract, trust by attestation

Foundry's privacy commitments are contractual: the Product Terms, the DPA, and Azure's audit evidence. The router documentation states "It does not store your prompts," and you rely on Microsoft's word plus its auditors. TrustedRouter's gateway runs inside hardware enclaves on three clouds — GCP Confidential Space, AWS Nitro Enclaves, and Azure Confidential Containers — with live attestation endpoints and a verifier script anyone can run from trust.trustedrouter.com. The attested build is bound to published source and release digests, and every line that touches your prompt is public: the gateway is source-available under BUSL-1.1, the SDKs Apache-2.0 or MIT.

We state the boundary plainly: attestation covers our gateway, never the downstream model providers. Providers are covered by cited policy and contract tiers — trustedrouter/zdr restricts routing to endpoints with contractual or policy zero-data-retention — while trustedrouter/e2e restricts routing to providers running confidential compute with provider-side end-to-end encryption, which are the providers' own mechanisms. Microsoft's structural commitment that Azure-hosted models keep prompts away from OpenAI entirely is documented in its data-privacy terms, and it is a property our attestation cannot give you.

Fees on different bases

Do not equate the numbers directly. Foundry's model router adds a markup of $0.14 per million input tokens on Global deployments ($0.154 Data Zone) on top of the underlying Azure model rates. As of August 2026, gpt-5.6-sol Global Standard is $5.00/M input and $30.00/M output; Data Zone runs about 10% higher, Priority Processing doubles it, Batch halves it, and Claude bills separately in CCUs. The public pricing page renders placeholders; the real numbers live in the Retail Prices API. TrustedRouter's prepaid text and embedding pricing is the provider's token price plus 5.5% with a $0.01/M floor, listed per model on /models and explained on /pricing; BYOK is supported, and our pricing page does not publish a separate BYOK fee. One is a flat per-input-token fee plus cloud list prices; the other is a percentage of provider price. Run your own model mix through both.

What we do not claim

Our repos are young: public since late April 2026, with benchmark history from June 2026. We publish a DPA and BAA but hold no published SOC 2 or HIPAA certification and no FedRAMP authorization — if those are hard requirements, Foundry wins today. Our catalog is 561 model entries (including our meta route ids) across roughly 49 providers as of August 16, 2026: far smaller in raw count than Foundry's 1,900+.

Moving a workload

OpenAI-shaped code moves by pointing the client at api.trustedrouter.com, swapping the key, and replacing Azure deployment names with catalog model ids; older code also drops the api-version query parameter. Claude-on-Foundry workloads already speak the Anthropic Messages shape, which we serve natively. What does not carry over: server-side state (Responses history, Assistants threads, stored completions, files and vector stores must be exported and rebuilt), Azure content-filter annotations in responses, and Azure-native machinery such as Foundry Agent Service tools, On Your Data, PTU reservations, and Azure Policy governance. Model-router deployments map to trustedrouter/auto plus provider preferences.

Migration shape

Start with one real request.

OpenAI-shaped Chat Completions and Responses code moves with a base URL, key, and deployment-name-to-model-id swap; Claude workloads keep the Anthropic Messages shape, which TrustedRouter serves natively. Server-side state (Assistants threads, stored completions, vector stores), Azure content-filter annotations, and Azure-native tooling like Foundry Agent Service, PTU reservations, and Azure Policy do not carry over.

Keep the first test small, stream the response, and compare output, latency, provider selection, and billed usage before moving production traffic.

Agent setup Run a small eval

TrustedRouter sideOpenAI SDK

from openai import OpenAI

client = OpenAI(
    base_url="https://api.trustedrouter.com/v1",
    api_key="sk-tr-v1-...",
)

with client.chat.completions.create(
    model="trustedrouter/zdr",
    messages=[{"role": "user", "content": "Reply PONG"}],
    stream=True,
) as response:
    for chunk in response:
        print(chunk.choices[0].delta.content or "", end="")

Official evidence

Sources checked August 16, 2026

Report a change

Data, privacy, and security for Foundry Models sold by Azurehttps://learn.microsoft.com/en-us/azure/ai-foundry/responsible-ai/openai/data-privacy
Modified abuse monitoring eligibility (Limited Access)https://learn.microsoft.com/en-us/azure/ai-foundry/responsible-ai/openai/limited-access
Foundry model router concepts (GA, cross-vendor pool)https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-router
What is Microsoft Foundry (catalog size, naming history)https://learn.microsoft.com/en-us/azure/foundry/what-is-foundry
Claude in Microsoft Foundry GA announcement (CCU billing)https://azure.microsoft.com/en-us/blog/claude-in-microsoft-foundry-is-now-generally-available/
Azure Retail Prices API: Model Router metershttps://learn.microsoft.com/en-us/rest/api/cost-management/retail-prices/azure-retail-prices
Azure services in FedRAMP audit scopehttps://learn.microsoft.com/en-us/azure/azure-government/compliance/azure-services-in-fedramp-auditscope
Docs commit removing the 30-day retention statement (2026-05-18)https://github.com/MicrosoftDocs/azure-ai-docs/commit/768d08d61e
TrustedRouter monthly benchmark reportsStable first-party route evidence
TrustedRouter live trust recordRelease and attestation evidence

All gateways

TrustedRouter vs Amazon BedrockCloud model platform
TrustedRouter vs Google Vertex AICloud model platform
TrustedRouter vs OpenRouterHosted model marketplace
TrustedRouter vs Vercel AI GatewayHosted AI gateway

Questions

Can we turn off Azure's abuse-monitoring storage?

Only through modified abuse monitoring, which Microsoft limits to customers managed by a Microsoft account team or in eligible programs; self-service customers cannot opt out. Once approved, the off state is verifiable: the resource JSON shows ContentLogging: "false", which is better verification of an opt-out than most providers offer. On TrustedRouter, no durable prompt or output logs on realtime inference is the default for every account, with no approval step.

Foundry has both GPT and Claude plus a GA router. Why add another gateway?

If your needs stop at those vendors on Azure, you may not need one — Foundry's dual-frontier catalog and cross-vendor router are real advantages and we say so plainly. TrustedRouter routes across roughly 49 providers as of August 2026, including confidential-compute hosts like Tinfoil and Phala, adds enforced privacy and region routes (trustedrouter/zdr, e2e, eu), and replaces policy trust in the gateway with hardware attestation you can check at trust.trustedrouter.com.

How do the fees actually compare?

They sit on different bases, so no single number answers it. Foundry charges a $0.14 per million input token router markup (Global; $0.154 Data Zone) on top of Azure model rates, with Data Zone about 10% higher, Priority Processing at 2x, and Batch at half price; Claude bills separately in Marketplace CCUs. TrustedRouter charges the provider's token price plus 5.5% with a $0.01/M floor, listed per model on /models. Run your own model mix through both; neither is uniformly cheaper.