Cloud model platform · Verified August 16, 2026

TrustedRouter vs Amazon Bedrock

Amazon Bedrock is AWS's managed model platform: ~120 serverless models across 18 providers, strong default privacy, and no gateway markup on your AWS bill. TrustedRouter is cloud-neutral, with a hardware-attested gateway and per-request privacy-tier routing across ~49 providers.

Compare details

8decision dimensions

8official sources

1base URL to try TR

Choose Amazon Bedrock when

Its operating model is the feature.

Choose Bedrock when your workloads, identity, and compliance evidence already live on AWS: IAM/SCP governance, PrivateLink, HIPAA eligibility (excluding the Fable and Mythos models), FedRAMP, spend that counts toward AWS commitments, and both Anthropic and OpenAI frontier models with no gateway markup.

Choose TrustedRouter when

Privacy needs evidence.

Choose TrustedRouter when you need one key across roughly 49 providers on any cloud, a gateway whose no-durable-prompt-logs behavior is checkable through live TEE attestation instead of policy alone, and per-request privacy routing (zdr, e2e, eu) that travels with the call.

DimensionAmazon BedrockTrustedRouter

DeploymentAWS regional managed service; IAM/SCP governance, PrivateLink, KMS, FIPS endpointsHosted control plane, source-available (BUSL-1.1), with an attested API path

API surfaceConverse/InvokeModel plus OpenAI-compatible and Anthropic APIs, split across 2 endpointsOpenAI Chat Completions and Responses plus Anthropic Messages

Model access~120 serverless models across 18 providers incl. GPT-5.x and Claude; 100+ Marketplace500+ models across ~50 providers: prepaid, BYOK, and direct

RoutingCross-region inference profiles (global ≈10% cheaper) and intelligent prompt routingProvider fallback plus auto, cheap, fast, free, ZDR, E2E, and EU routes with composable privacy and jurisdiction preferences

ObservabilityCloudWatch/CloudTrail; prompt logging opt-in, default off, absent on mantle endpointMetadata analytics and opt-in external broadcast

Prompt contentNo prompt storage by default; GPT-5.x flagged traffic and Fable/Mythos retained ≤30dNo durable prompt or output logs on realtime inference; batch is opt-in encrypted retention

VerificationIAM, audit logs, SOC 1/2/3, FedRAMP, HIPAA-eligible; policy trust, no runtime attestationLive gateway attestation on three clouds, bound to published source and release evidence

BillingAWS metering at listed model prices, no markup; batch/flex −50%, priority +75%Prepaid at provider price + 5.5% ($0.01/M floor), or BYOK

What Amazon Bedrock actually is

Amazon Bedrock is AWS's fully managed foundation-model service. AWS copies each provider's model into AWS-controlled deployment accounts — one per provider per region — and the providers have no access to those accounts, so they cannot see customer prompts or completions. As of August 2026 the serverless catalog is roughly 120 models across 18 providers, including Anthropic's Claude line through Fable 5 and OpenAI's proprietary GPT-5.x family (GA June and July 2026), with 100+ more models available through the Bedrock Marketplace. Bedrock is currently the only place other than the first parties where Anthropic and OpenAI frontier models sit behind one billing relationship.

There are two inference endpoints, and the split matters when you write code. bedrock-runtime carries the AWS-native Converse and InvokeModel APIs, OpenAI-compatible Chat Completions and Responses, Anthropic Messages, Guardrails, and cross-region inference profiles. bedrock-mantle carries the OpenAI-compatible and Anthropic APIs with server-side tools, background inference, and Projects, and lacks Guardrails, cross-region profiles, and model invocation logging. Authentication is IAM SigV4 or plain API keys that work directly with the OpenAI SDK.

Where Bedrock is the right choice

Bedrock's default privacy posture is exceptionally strong for a hosted model platform. Model inputs and outputs are not stored by default, no AWS operator can read them, customer content is not used to train base models, and the 2026 Data Retention API adds a none mode, enforceable org-wide through SCPs, in which any model that requires retention becomes unavailable rather than retaining. Compliance depth is real: HIPAA-eligible (excluding the Fable and Mythos models), FedRAMP Class C and Class D including GovCloud, SOC 1/2/3, FIPS-validated endpoints, and PrivateLink.

Pricing is a genuine strength too. There is no gateway markup: you pay AWS's listed per-token price, metered on your AWS bill, counting toward committed spend. AWS passed OpenAI's July 30, 2026 cuts through — GPT-5.6 Luna dropped 80% to $0.20 input / $1.20 output per million tokens — and batch and flex tiers each take 50% off, with global cross-region routing about 10% cheaper than geographic. If your infrastructure, identity, and compliance program already live on AWS, Bedrock is a strong default and this page will not talk you out of it.

Three differences that decide it

Verification kind. Bedrock's no-storage posture is policy backed by audits: AWS documentation, SOC reports, contract terms. TrustedRouter makes the equivalent claim mechanically checkable: api.trustedrouter.com runs inside hardware enclaves on three clouds (GCP Confidential Space, AWS Nitro Enclaves, Azure Confidential Containers), and anyone can verify that the running build matches the published source and release digests from trust.trustedrouter.com. Realtime inference does not write prompt or output content to durable storage, and every line of gateway code that touches your prompt is public.

Carve-outs versus routes. Bedrock's zero-retention default now has flagship exceptions: classifier-flagged GPT-5.x traffic is retained up to 30 days, and Claude Fable 5 and Mythos 5 require retention with data sharing to Anthropic and potential human review — AWS's launch blog says plainly that once you opt in to data retention, "your data will leave AWS's data and security boundary." Bedrock manages this with account-level retention modes, and eligible customers can request full zero-data-retention for retention-requiring models through their AWS account team. TrustedRouter manages it with routing: trustedrouter/zdr restricts a request to providers with a cited zero-data-retention posture, trustedrouter/e2e restricts to confidential-compute providers with end-to-end encryption (tinfoil, phala), and trustedrouter/eu pins the EU provider set. Privacy floors compose with any request, so the constraint travels with the call rather than the account.

Cloud neutrality, with the fee stated. Bedrock is one cloud: IAM identity, per-region model availability, and no Google Gemini (only open-weight Gemma). TrustedRouter routes 550+ model routes across roughly 49 providers (as of August 2026) with one key, including Google routes. The pricing bases differ, so here are both: Bedrock adds no fee over its listed model prices; TrustedRouter prepaid bills the provider's token price plus 5.5%, with a $0.01 per million token floor (pricing). Where Bedrock's listed price matches the provider's own — AWS matched OpenAI's first-party cuts, for example — Bedrock is cheaper by our 5.5% fee. Base prices are not identical everywhere: Bedrock lists legacy Claude 3.5 Sonnet at a 2x extended-access surcharge, so compare the exact SKUs you run. BYOK is supported; we do not publish a separate BYOK fee.

What we do not claim

Our attestation covers the gateway and stops there. Downstream model providers are covered by cited policy tiers, except the E2E routes (tinfoil, phala), where the provider's own confidential-compute and encryption mechanisms apply.
We publish no SOC 2 or HIPAA certification. If either is a hard requirement today, Bedrock wins this comparison outright.
Our repos are young: public since late April 2026, with benchmark history from June 2026. AWS's operating track record is measured in decades.
The gateway and control plane are source-available under BUSL-1.1, converting to Apache-2.0 four years after each release; the SDKs are Apache-2.0 or MIT. We do not call the platform open source.

Migration reality

If your app already uses Bedrock's OpenAI-compatible Chat Completions or Responses APIs, moving is mechanical: change the base URL and API key, then remap Bedrock model ids (such as anthropic.claude-... or inference-profile ARNs) to catalog names on /models. Apps built on Converse or InvokeModel need a request-shape rewrite first; those are AWS-native shapes no gateway speaks.

What does not carry over: IAM and SCP enforcement, Guardrails, Knowledge Bases and Agents, PrivateLink, CloudWatch and CloudTrail wiring, and spend that counted toward AWS commitments all stay behind. Stateful Responses conversations created with store=true are pinned to the AWS region that served them, so export or reset that state before cutover. Start with one streamed request, compare output, latency, and billed usage, and read /security and the status page before routing anything sensitive.

Migration shape

Start with one real request.

Apps on Bedrock's OpenAI-compatible endpoints move with a base URL, key, and model-id change. Converse or InvokeModel apps need a request-shape rewrite first. IAM/SCP enforcement, Guardrails, Knowledge Bases, PrivateLink, CloudWatch wiring, and AWS-commitment spend do not carry over; Responses state stored with store=true stays in the AWS serving region.

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

Bedrock data retention modeshttps://docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html
Bedrock abuse detection and default no-storagehttps://docs.aws.amazon.com/bedrock/latest/userguide/abuse-detection.html
Bedrock runtime vs mantle endpoint comparisonhttps://docs.aws.amazon.com/bedrock/latest/userguide/endpoints.html
Bedrock models at a glancehttps://docs.aws.amazon.com/bedrock/latest/userguide/model-cards.html
Amazon Bedrock pricinghttps://aws.amazon.com/bedrock/pricing/
AWS HIPAA eligible services listhttps://aws.amazon.com/compliance/hipaa-eligible-services-reference/
GPT-5.6 Luna/Terra price reduction (AWS What's New)https://aws.amazon.com/about-aws/whats-new/2026/07/openai-gpt-terra-luna-pricing-bedrock/
Claude Fable 5 on AWS launch bloghttps://aws.amazon.com/blogs/aws/anthropic-claude-fable-5-on-aws-mythos-class-capabilities-with-built-in-safeguards-now-available/
TrustedRouter monthly benchmark reportsStable first-party route evidence
TrustedRouter live trust recordRelease and attestation evidence

All gateways

TrustedRouter vs Microsoft Azure AI FoundryCloud model platform
TrustedRouter vs Google Vertex AICloud model platform
TrustedRouter vs OpenRouterHosted model marketplace
TrustedRouter vs Vercel AI GatewayHosted AI gateway

Questions

Bedrock already does not store prompts by default. What does attestation add?

Bedrock's default posture is excellent and we say so: no input/output storage by default, zero operator access, and model providers architecturally cut off from prompts. That posture is enforced by AWS policy and verified by auditors. TrustedRouter's no-durable-prompt-logs claim is verified differently: the gateway runs in enclaves on three clouds, and anyone can check the running build against published source from trust.trustedrouter.com without waiting on an audit cycle. The attestation covers our gateway only; it does not extend to downstream model providers.

Which catalog is bigger, and what is each one missing?

Bedrock lists ~120 serverless models across 18 providers plus 100+ Marketplace models, and it is the only non-first-party host of both Anthropic and OpenAI frontier models; it carries no Google Gemini, only open-weight Gemma. TrustedRouter lists 550+ model routes across ~49 providers (as of August 2026), including Google routes. Coverage gaps exist on both sides, so check /models for the specific SKUs you need before committing.

What breaks when we move off Bedrock?

OpenAI-compatible call sites just re-point; Converse and InvokeModel call sites need rewriting to OpenAI-compatible shapes. You lose IAM/SCP enforcement, Guardrails, Knowledge Bases and Agents, PrivateLink, CloudWatch/CloudTrail wiring, and spend that counted toward AWS commitments, so plan replacements before cutover. Responses state stored with store=true stays in the AWS serving region; export it first. Our only step-by-step migration doc today covers OpenRouter, not Bedrock.