Vendor head-to-head · 1 May 2026

AWS Bedrock vs Azure OpenAI / Foundry: enterprise agent platforms

In 2026 most enterprise agent deployments run on AWS Bedrock or Azure OpenAI / AI Foundry rather than direct vendor APIs. The hyperscaler-mediated path simplifies procurement (one MSA covers the model + infrastructure), but it constrains the model selection and changes the residency, security, and audit posture in ways that the direct path does not. This comparison is for IT leaders selecting the agent-platform layer rather than the foundation model itself. The decision often follows the in-house cloud commitment, but the platform features, model breadth, and contractual primitives diverge enough that the cloud commitment alone shouldn't decide it.

Who this is for

· Enterprise architects choosing the agent-platform layer for 2026 production deployment
· Procurement leads negotiating committed-spend agreements with AWS or Azure
· CISOs evaluating shared-responsibility model boundaries for hosted AI inference

Side A

AWS Bedrock ↗

AWS-managed agent platform offering frontier and open-weight models from Anthropic, Meta, Mistral, Cohere, AI21, Stability, and Amazon. Bedrock Agents for orchestration; Knowledge Bases for RAG; Guardrails for content controls.

PricingFrom $3/M input tokens (Claude 3.5 Sonnet on Bedrock)· as of 1 May 2026source ↗

Side B

Azure OpenAI / AI Foundry ↗

Microsoft-managed agent platform centred on OpenAI models (GPT-4, GPT-5, o-series) plus selected partner models in AI Foundry. Azure AI Agent Service for orchestration; Microsoft Agent Framework as the agent SDK.

PricingFrom $2.50/M input tokens (GPT-4o on Azure)· as of 1 May 2026source ↗

Feature matrix

Dimension	AWS Bedrock	Azure OpenAI / AI Foundry
Model breadthsource ↗	Anthropic Claude family, Meta Llama, Mistral, Cohere, AI21, Amazon Nova, Stability AI	OpenAI GPT/o-series (primary); curated partner models (Llama, Mistral, Phi, Cohere, DeepSeek) in AI Foundry
EU residency posturesource ↗	Frankfurt (eu-central-1), Ireland (eu-west-1), Paris, London regions; AWS European Sovereign Cloud rolling out 2025-2026	Sweden Central, France Central, Switzerland North, Germany West Central; EU Data Boundary covers most processing
Agent orchestration primitivesource ↗	Bedrock Agents (managed); Step Functions for code-first orchestration	Azure AI Agent Service (managed); Microsoft Agent Framework SDK; Logic Apps for workflow
RAG / retrieval primitivesource ↗	Bedrock Knowledge Bases (managed); OpenSearch / Aurora pgvector for self-managed	Azure AI Search (managed); Cosmos DB pgvector + Postgres for self-managed
Guardrails / content filtersource ↗	Bedrock Guardrails (deny topics, PII redaction, prompt-injection filter, denied-words filter)	Azure AI Content Safety (built-in moderation, prompt shields, groundedness detection)
Audit substratesource ↗	CloudTrail + Bedrock model invocation logs; CloudWatch metrics + alarms	Azure Monitor + Application Insights; Microsoft Purview audit integration
Compliance attestations (2026)source ↗	SOC 1/2/3, ISO 27001/27017/27018/27701/42001, HIPAA, PCI DSS, FedRAMP High, IRAP, C5	SOC 1/2/3, ISO 27001/27017/27018/27701/42001, HIPAA, PCI DSS, FedRAMP High, IRAP, C5
Marketplace / commitment-spend integrationsource ↗	AWS Marketplace; Bedrock charges count against EDP / PPA commitments	Azure Marketplace; Azure OpenAI charges count against MACC / EA commitments
Hosted Anthropic Claude availabilitysource ↗	Yes — Sonnet, Opus, Haiku (Anthropic-on-Bedrock is the most common enterprise distribution path for Claude)	No — Anthropic models not available on Azure OpenAI / Foundry
Hosted OpenAI availabilitysource ↗	No — OpenAI models not available on Bedrock	Yes — GPT-4, GPT-4o, GPT-5, o-series (Azure OpenAI is the enterprise distribution channel)

What our claim ledger says about each

When to choose which

Choose AWS Bedrock

Pick AWS Bedrock when the in-house stack is AWS-resident, when the deployment needs Anthropic Claude (Bedrock is the enterprise distribution channel for Claude), or when model breadth across vendors matters. Stronger fit when the architecture wants vendor-substitution flexibility within a single platform — Bedrock's model-vendor diversity is materially wider than Foundry's.

Choose Azure OpenAI / AI Foundry

Pick Azure OpenAI / Foundry when the in-house stack is Azure-resident, when the deployment needs OpenAI's frontier models (GPT-5, o-series), or when the workload integrates with Microsoft Graph / Entra ID natively. Stronger fit when foundation-model commitment to OpenAI is already in place and the vendor lock-in is a deliberate procurement decision.

Open questions we're tracking

AM-108· next review +57dAgentic-AI data-residency requirements are not cleanly inherited from existing GDPR cross-border transfer practice. Agent context windows, retrieval indexes, and reasoning traces all create new categories of personal-data processing that have to be located, documented, and (for high-risk Annex III deployments) data-resident inside the EEA before EU AI Act Article 16 enforcement opens on 2 August 2026. The deployment topology has to shift to single-region EEA-resident for high-risk systems; hub-and-spoke remains defensible for general-purpose deployments under documented GDPR Chapter V transfer mechanisms.