GILD verdict
Choose by model family first, then by where your cloud agreement lives.
Teams that run both, such as DoorDash and Fin, buy OpenAI models through Azure and Claude through Bedrock: Foundry for GPT, with provisioned throughput in Global, Data Zone or regional deployments, and Bedrock for Claude, with EU Geo profiles and AWS as processor. Choose Foundry for Claude only when you need Opus 5.5 fast mode, which runs on its Azure-hosted deployments and not on Bedrock. Requesty measured 96% request success on Bedrock against 78% on Azure in April 2026, so test both under your own load before committing spend.
Teams that run both clouds split them by model family: DoorDash backs its OpenAI endpoint with Azure and its Claude endpoint with Bedrock (InfoQ, Oct 3, 2026), and Fin serves OpenAI models through Azure and Anthropic models through Bedrock, Vertex and Anthropic (Aug 8, 2025). Measured reliability points the same way: Requesty's gateway saw 96% of Bedrock requests complete in April 2026, against 78% on Azure. List prices for Claude Opus 5.5 match on Global routing, so at GILD's workload both cost US$4,800 a month; only Foundry's Azure-hosted deployments sell Opus 5.5 fast mode, at twice that.
What teams report
Run both 2
DoorDash
DoorDash's LLM Gateway uses Azure as the fallback for its OpenAI endpoint and Bedrock as the fallback for its Claude API endpoint, so each model family has a second cloud route (InfoQ talk, Oct 3, 2026).
“We actually had to go understand what the fallback should be, what kind of routing should be. At one point, I remember we had OpenAI endpoint, but we also had a fallback on Azure endpoint. Or, we had Claude model API endpoint, then we had a fallback on Bedrock, so that the services can rely on us from the reliability point of view. We went in that direction. We quickly realized that as the usage is going up, then we need a way to track, because now finance got involved. Like, you guys have one API key for multiple teams, how do I track who is using what? We quickly realized we need workspaces where we have cost attribution. We can actually drill down on who is using how much. We can attribute cost back to those departments. That's what basically we did with LLM Gateway.”
Siddharth Kodwani, Software Engineer, GenAI Platform, DoorDashInfoQ, , recording Fin
Fin serves OpenAI models through Azure or OpenAI and Anthropic models through Bedrock, Vertex or Anthropic, and shifted traffic between providers during the OpenAI outage of June 10, 2025 (Aug 8, 2025).
“For example, Anthropic models can be served by AWS Bedrock, GCP Vertex, or Anthropic’s own infrastructure; OpenAI models are served by either Azure or OpenAI itself.”
Fin AI research team, Research team, FinFin AI Research, , write-up
Choosing between Amazon Bedrock and Microsoft Foundry
Choose Amazon Bedrock if
- Your workloads, data, and IAM already run in AWS.
- You want Claude served on AWS infrastructure under AWS's data terms, alongside Amazon Nova and open-weight models, with Geo profiles that keep Opus 5.5 inside the EU.
- You want AgentCore to run agents built with any framework, with Gateway, Identity, and Policy around tool calls.
Choose Microsoft Foundry if
- Your identity and data sit in Azure and Microsoft Entra ID.
- You buy OpenAI models sold directly by Azure, with provisioned throughput units in Global, Data Zone, or regional deployments, the split DoorDash and Fin use for OpenAI traffic.
- You want Claude hosted on Azure or on Anthropic infrastructure, with tools such as file search and code execution, on a paid Azure subscription, or Opus 5.5 fast mode on Azure-hosted deployments.
Side by side
| Dimension | Amazon Bedrock | Microsoft Foundry |
|---|---|---|
| Monthly bill at one workloadGILD's calculation from published rates, fetched Oct 28, 2026: 2,000 tasks a day for 30 days, each with 10,000 input tokens and 2,000 output tokens and no caching, at Opus 5.5's US$4 input and US$20 output per million tokens, with 10% added for Geo or Data Zone routing and fast mode at US$8 and US$40. Taxes, committed-spend discounts and negotiated rates are excluded. | US$4,800 a month on Claude Opus 5.5 on Global routing; US$5,280 on Geo routing. No fast mode322 | US$4,800 a month on Claude Opus 5.5 at the same list price; US$5,280 in the US Data Zone. US$9,600 in fast mode on Azure-hosted deployments2930 |
| Claude Opus 5.5 fast mode | Not offered; the model card lists the Standard tier only22 | Offered on Azure-hosted deployments as a separate claude-opus-5-5-fast model at US$8 input and US$40 output; not on Anthropic-hosted deployments29 |
| Strongest model family | Claude: EU and other Geo profiles for Opus 5.5, and AWS as the data processor22 | GPT: OpenAI models sold directly by Azure, with provisioned throughput units in Global, Data Zone or regional deployments24 |
| Model catalog | More than 100 models from Amazon, Anthropic, OpenAI, DeepSeek, Meta, Mistral AI, Cohere, AI21 Labs, and other providers | More than 10,000 models from Microsoft, OpenAI, Anthropic, Meta, and others; Azure OpenAI models are sold directly by Azure |
| Claude | Served on Bedrock's own inference engine with zero operator access, at Anthropic's list prices on Global routing, through the Messages, Converse, and Invoke APIs | Two versions, hosted on Azure or on Anthropic infrastructure; requires a paid Azure subscription, and credit-only, sponsored, and Cloud Solution Provider subscriptions are not supported |
| Pay-per-token options | Standard, Priority (75% premium), and Flex (50% off) where a model supports them; Claude Opus 5.5 Geo routing adds 10% over Global | Global Standard, Data Zone Standard, and regional Standard; Global Standard launches first at the lowest price |
| Batch | 50% below on-demand for supported models, as jobs reading from and writing to Amazon S3 | Global Batch and Data Zone Batch at a 50% discount, with a 24-hour window |
| Reserved capacity | Reserved tier for 1 or 3 months from 100,000 input TPM, and Provisioned Throughput in hourly Model Units with no commitment, 1-month, or 6-month terms | Provisioned throughput units in Global, Data Zone, and regional deployments |
| Agents | AgentCore runs agents from any framework with Gateway, Memory, Identity, Policy, and Observability; Agents Classic closed to new customers; Managed Agents by OpenAI not yet generally available | Foundry Agent Service: prompt agents, hosted agents running your own code, and toolboxes with web search, file search, code interpreter, and MCP servers |
| Data handling | Inputs and outputs are not shared with model providers or used to improve base models | Prompts and completions are not available to OpenAI or other model providers and not used to train models; abuse monitoring applies, with an application to modify it |
| Data residency | In-Region, Geo cross-Region, or Global routing, set per model | Global, Data Zone (EU, US, APAC), or regional deployments within an Azure geography |
| Identity and governance | IAM and Service Control Policies, PrivateLink, CloudTrail | Microsoft Entra ID, role-based access control, network isolation, and Azure Policy |
| Measured success rate | 96% of requests completed without an upstream error on Requesty's gateway in April 20261 | 78% on Azure in the same data, the lowest of the major routes Requesty measured1 |
What practitioners report, side by side
Each product's figures come from its own practitioners, for the group you pick.
Overall signal
All practitioners rate Amazon Bedrock +3, 29 points above Microsoft Foundry at −26. Most of the gap is in complaints, with 20 of Microsoft Foundry's 36 sources negative against 41 of Amazon Bedrock's 108.
The score counts the authors from all practitioners that take a side, one vote each with their latest view, and is how far positive votes lead negative ones, per 100 votes, so Amazon Bedrock's +3 rests on votes split 41 positive, 22 mixed and 38 negative and Microsoft Foundry's −26 on 9, 7 and 18.
A lead of a fifth of the votes reads positive and a lead of half strong positive, with the same lines below zero. The dots show each source, positive, mixed, negative. Run a pilot on your own prompts to measure output quality and cost per task.
Footprint on the AI Application Platform stack
Both cover the same 9 layers at the same depth. Coverage is each product's placement on this stack, as its product page shows it. Core is a job the product owns; Supports and Touches are jobs it helps with. Coverage says nothing about how well a product does them.
| Stage | Amazon Bedrock | Microsoft Foundry |
|---|---|---|
| Model2 layers | Core 1, Supports 1 | Core 1, Supports 1 |
| Build2 layers | Supports 2 | Supports 2 |
| Data1 layer | Supports 1 | Supports 1 |
| Evaluate1 layer | Supports 1 | Supports 1 |
| Operate3 layers | Core 1, Supports 2 | Core 1, Supports 2 |
Show all 9 layers
| Layer | Amazon Bedrock | Microsoft Foundry |
|---|---|---|
| Model | ||
| Model APIs | Core | Core |
| Model gateway | Supports | Supports |
| Build | ||
| Orchestration | Supports | Supports |
| Tool use | Supports | Supports |
| Data | ||
| Retrieval | Supports | Supports |
| Evaluate | ||
| Evaluation | Supports | Supports |
| Operate | ||
| Observability | Supports | Supports |
| Cloud deployment | Core | Core |
| Guardrails | Supports | Supports |
What practitioners praise
Among all practitioners, Amazon Bedrock alone has seven praise themes that reach 5 sources from 3 authors. Model Choice and AWS Integration lead, at 17 of 27 and 11 of 11. Both carry Developer Experience. It is praised in 0 of the 7 Amazon Bedrock sources that discuss it, against 3 of 10 for Microsoft Foundry. Microsoft Foundry alone has three praise themes that reach 5 sources from 3 authors. Model Catalog & Choice and Contracts, Data Terms & Identity lead, at 11 of 16 and 8 of 10.
| Praised for | Amazon Bedrock | Microsoft Foundry |
|---|---|---|
| Developer Experience | 0 of 7 | 3 of 10 |
| Model Choice | 17 of 27 | Not a theme in its sources |
| AWS Integration | 11 of 11 | Not a theme in its sources |
| Cost Efficiency | 10 of 12 | Not a theme in its sources |
| Speed | 6 of 7 | Not a theme in its sources |
| Model Quality | 6 of 9 | Not a theme in its sources |
| Stability | 5 of 8 | Not a theme in its sources |
| Managed Agents, RAG & Guardrails | 4 of 10 | Not a theme in its sources |
| Model Catalog & Choice | Not a theme in its sources | 11 of 16 |
| Contracts, Data Terms & Identity | Not a theme in its sources | 8 of 10 |
| Cost | Not a theme in its sources | 3 of 5 |
Each count is the sources that praise a theme, or raise it as a concern, out of those that discuss it. A row pairs the two products when both carry a theme under the same or an equivalent name.
A count ending in "so far" has under 5 sources from 3 authors among all practitioners. "Not a theme in its sources" means GILD found no theme of that name in that product's sources.
What practitioners raise as concerns
Among all practitioners, Amazon Bedrock alone has nine concern themes that reach 5 sources from 3 authors. Quotas, Capacity & Latency and Setup & Model Access lead, at 20 of 25 and 12 of 13. Microsoft Foundry alone has six concern themes that reach 5 sources from 3 authors. Agent Service & Tool Failures and Missing Models & Features lead, at 9 of 9 and 8 of 8.
| Raised as a concern | Amazon Bedrock | Microsoft Foundry |
|---|---|---|
| Quotas, Capacity & Latency | 20 of 25 | Not a theme in its sources |
| Setup & Model Access | 12 of 13 | Not a theme in its sources |
| Tool Calling Failures | 10 of 10 | Not a theme in its sources |
| Cost Surprises | 9 of 11 | Not a theme in its sources |
| Lag Behind Vendor APIs | 9 of 11 | Not a theme in its sources |
| Inconsistent Results | 8 of 8 | Not a theme in its sources |
| Data Terms & Retention | 8 of 11 | Not a theme in its sources |
| Latency | 6 of 6 | Not a theme in its sources |
| Regional Availability | 6 of 11 | Not a theme in its sources |
| Agent Service & Tool Failures | Not a theme in its sources | 9 of 9 |
| Missing Models & Features | Not a theme in its sources | 8 of 8 |
| Reliability & Latency | Not a theme in its sources | 7 of 7 |
| Portal, Naming & Setup | Not a theme in its sources | 6 of 7 |
| Quotas, Capacity & Cost | Not a theme in its sources | 5 of 5 |
| Data Terms & Model Governance | Not a theme in its sources | 3 of 6 |
32 sources
- Requesty: Operational metrics per provider, April 2026 (May 9, 2026)
- Amazon Bedrock User Guide: Overview
- Amazon Bedrock pricing
- AWS model card: Anthropic Claude Opus 5.5
- AWS model card: Amazon Nova Pro
- AWS model card: Meta Llama 4 Scout 17B
- AWS model card: Mistral Large 3
- AWS model card: DeepSeek V3.2
- AWS model card: OpenAI gpt-oss-120b
- AWS model card: Cohere Command R
- AWS model card: AI21 Jamba 1.5 Mini
- AWS price list: metered model rates
- AWS price list: foundation model rates
- Amazon Bedrock User Guide: Service tiers
- Amazon Bedrock User Guide: Batch inference
- Amazon Bedrock User Guide: Provisioned Throughput
- Amazon Bedrock FAQs
- Amazon Bedrock AgentCore
- Amazon Bedrock User Guide: Agents
- Amazon Bedrock User Guide: Data retention
- Amazon Bedrock User Guide: Abuse detection
- Amazon Bedrock User Guide: Claude Opus 5.5 model card
- Microsoft Learn: What is Microsoft Foundry?
- Microsoft Learn: Deployment types for Microsoft Foundry Models
- Microsoft Learn: Data, privacy, and security for Foundry Models
- Microsoft Learn: Foundry Models from partners and community
- Microsoft Learn: Agents in Microsoft Foundry
- Microsoft Foundry pricing
- Claude API docs: Fast mode, where Opus 5.5 fast mode runs (as of Oct 28, 2026)
- OpenRouter: Claude Opus 5.5 endpoints, including Bedrock and Azure (Oct 28, 2026)
- InfoQ: Building a GenAI Platform at DoorDash, talk by Siddharth Kodwani and Swaroop Chitlur (Oct 3, 2026)
- Fin AI Research: Running a reliable service over unreliable parts (Aug 8, 2025)