What teams building AI features on Microsoft Foundry report, with every finding tied to its evidence.
Read the findingsPractitioner sentiment, all practitioners
−26 Negative from 36 sources that take a side
How this figure is madeFrom 34 authors in 35 discussions. Run a pilot on your own prompts to measure output quality and cost per task.
Jan 2025 to Oct 2026.
45 sources from practitioners, by month and sentiment.
Accounts lean negative. The model catalog and the enterprise data terms draw most of the praise, and contract and data terms are the most common reason given for choosing Foundry. Agent Service failures and missing models or features are the most reported concerns, followed by reliability and a confusing portal.
Positive, mixed and negative sources
The score counts them as 9 positive, 7 mixed and 18 negative votes, one per author. See How the numbers work.
See the evidenceBelow: 4 praise themes, 6 concerns, 1 point of agreement and 1 disagreement.
Counted for All practitioners, from 45 sources. Change group
Every capability practitioners describe positively, ranked by how many sources say so. Open one to read who said it.
Counted for All practitioners.
Positive, 11
Positive
kasey_junk, developer
DeepInfer: High-performance inference framework for DeepSeek modelsHacker News comment
The author can deploy open-weight models in Foundry with one click, already approved by the company's compliance and procurement teams. Data contracts are why the author expects cloud platforms to win enterprise model business.
Positive
jiggawatts, consultant
Using Opus 5.5 to discover a new eyewitness record of the dodoHacker News comment
The author says Foundry has every frontier model except Google's, so a Foundry subscription plus a Google one covers a workflow that compares several models on the same input.
Positive
taspeotis, developer
Grok 4.7Hacker News comment
The author deployed DeepSeek V4.1 Flash on the team's Foundry resource, ran it through Pi, and liked its tone.
No opinion stated
jiocrag, developer
GPT-6 AstraHacker News comment
The author notes that the latest OpenAI models have kept arriving in Foundry.
No opinion stated
pkaye, developer
Kimi K2.7 Code is generally available in GitHub CopilotHacker News comment
The author notes that the DeepSeek V4 models were already in Foundry.
Positive
fg137, developer
IBM CEO says open weight models are now good and cheap on-premHacker News comment
The author argues that companies should rent inference and skip buying GPUs, because Foundry and Amazon Bedrock already offer many open and closed models with zero data retention.
Positive
cdnsteve, developer
OpenCode 2.0Hacker News comment
The author's organization of more than 100 developers hosts its own models in Foundry and points Claude Code at them. Developers need no API keys because they sign in with az login.
Mixed
mvATM99, developer
Kimi K2.7 Code is generally available in GitHub CopilotHacker News comment
Adding a Foundry model as a GitHub Copilot custom model failed for the author, and the path is poorly documented and unstable. VS Code's bring-your-own-key route worked, so the author now uses DeepSeek V4 Flash with Copilot.
Positive
dandinu, developer
I built a WhatsApp clone of myself; a friend agreed on a password with itHacker News comment
The author built a WhatsApp bot that answers as the author, with Claude Sonnet 4.6 through Foundry generating each reply, in about two hours of work. Runtime cost is about US$0.005 per reply.
Positive
consumer451, developer building products for clients
Cybersecurity researchers aren't happy about the guardrails on Anthropic's FableHacker News comment
The author deploys Claude models on Azure because clients already trust their Microsoft contracts, and says the Anthropic models make the products much better. The author asks how data terms for newer frontier models affect that trust.
Mixed
zubiaur, developer
Microsoft has a problem: lack of demand for its AI productsHacker News comment
The author calls Azure's AI services decent and access to OpenAI models through Foundry a real differentiator, but says rate limits are a problem. The author worries the Foundry rebranding signals marketing taking over the services.
Mixed or negative, 5
Mixed
Rajesh Beri, enterprise AI practitioner and author of THE DAILY BRIEF
Bedrock vs Vertex AI vs Foundry: Your Cloud Commit DecidesNewsletter analysis, THE DAILY BRIEF
The author rates Foundry the strongest of the three platforms for GPT models, with the most flexible provisioned throughput, and the weakest for Claude: no Claude provisioned throughput, no EU data zone, and Anthropic as the data processor. At list prices the platforms differ by at most about US$300 a month on a fixed Claude workload, so existing committed spend should decide.
Mixed
chris_money202, developer
GPT 5.6 Sol Still Fails on Basic Code ReviewHacker News comment
Checking the Foundry catalog, the author found that the separate reasoning variants of a model seemed to be gone, leaving what may be one model.
Negative
cdnsteve, developer
DeepSeek v4.1 FlashHacker News comment
The author finds Foundry slow to add new models and missing many open-weight ones.
Negative
Julian Goldsmith (juliangoldsmith), developer
The LLM Lobotomy?Hacker News comment
Foundry still lacked tool calling for Llama models more than six months after Microsoft said it was working on it, and small models were consistently slow. Provisioned throughput would fix speed but started at about $10k a month for GPT models and about $72k for DeepSeek, against a few dollars a month of current use.
Negative
vemonet, developer
DeepSeek R1 Is Now Available on Azure AI Foundry and GitHubHacker News comment
DeepSeek R1 on Foundry took 80 seconds to answer a question GPT-4o answered in 7, under rate limits and a 4,000-token input cap at launch. The author also found errors in the inference SDK's documentation and choppy streaming from Mistral Large.
Counted for All practitioners.
Positive, 8
Positive
fg137, developer
IBM CEO says open weight models are now good and cheap on-premHacker News comment
The author argues that companies should rent inference and skip buying GPUs, because Foundry and Amazon Bedrock already offer many open and closed models with zero data retention.
Positive
kasey_junk, developer
DeepInfer: High-performance inference framework for DeepSeek modelsHacker News comment
The author can deploy open-weight models in Foundry with one click, already approved by the company's compliance and procurement teams. Data contracts are why the author expects cloud platforms to win enterprise model business.
Positive
cdnsteve, developer
OpenCode 2.0Hacker News comment
The author's organization of more than 100 developers hosts its own models in Foundry and points Claude Code at them. Developers need no API keys because they sign in with az login.
Positive
edot, developer
Navier-Stokes - Tristan Buckmaster [pdf]Hacker News comment
For teams that do not trust the model vendors with their data, the author points to Foundry or Amazon Bedrock, which earn money from renting infrastructure and so have a stronger reason not to train on customer data.
Positive
johnbarron, developer
OpenRouter is joining StripeHacker News comment
The author says corporate compliance teams reject third-party model routers and recommends coding enterprise apps with fallback loops across cloud platforms such as Foundry and Amazon Bedrock, which offer private endpoints and legal indemnification.
Positive
consumer451, developer building products for clients
Cybersecurity researchers aren't happy about the guardrails on Anthropic's FableHacker News comment
The author deploys Claude models on Azure because clients already trust their Microsoft contracts, and says the Anthropic models make the products much better. The author asks how data terms for newer frontier models affect that trust.
No opinion stated
827a, developer
Anthropic S-1Hacker News comment
The author expects enterprises to keep buying from cloud platforms, such as Foundry with zero data retention, before they would run their own GPU clusters.
Mixed
Rajesh Beri, enterprise AI practitioner and author of THE DAILY BRIEF
Bedrock vs Vertex AI vs Foundry: Your Cloud Commit DecidesNewsletter analysis, THE DAILY BRIEF
The author rates Foundry the strongest of the three platforms for GPT models, with the most flexible provisioned throughput, and the weakest for Claude: no Claude provisioned throughput, no EU data zone, and Anthropic as the data processor. At list prices the platforms differ by at most about US$300 a month on a fixed Claude workload, so existing committed spend should decide.
Mixed or negative, 2
Negative
jiggawatts, consultant
GPT-6 Astra on OpenRouterHacker News comment
The author says zero data retention is the main reason customers pick Foundry over OpenAI's own API, yet the model list makes it hard to see which models carry it. There is no simple way to block models without it, so each customer vets and approves models one by one.
Negative
arielcostas, developer
Getting a Gemini API key is an exercise in frustrationHacker News comment
The author found the setup confusing: a choice between Azure OpenAI and Microsoft Foundry resources, a separate portal with an old and a new Foundry experience, Foundry and Foundry Project resource types, and a single API key unless Entra ID is set up.
GILD's fact checks, which don't count as sources
Checked Oct 28, 2026.Microsoft's security baseline for Foundry lists Azure AD (Microsoft Entra ID) authentication for data plane access as supported and on by default, with access managed through Azure role-based access control and predefined roles assigned to users in Entra ID. Conditional Access policies are supported, and the customer configures them. Azure security baseline for Microsoft Foundry
Checked Oct 28, 2026.Foundry supports two authentication methods, Microsoft Entra ID and API keys. Microsoft recommends Entra ID for production because it enables conditional access, multifactor authentication, managed identities and least-privilege role-based access control. Authentication and authorization in Microsoft Foundry
Open them in the recordCounted for All practitioners.
Positive, 3
Positive
dandinu, developer
I built a WhatsApp clone of myself; a friend agreed on a password with itHacker News comment
The author built a WhatsApp bot that answers as the author, with Claude Sonnet 4.6 through Foundry generating each reply, in about two hours of work. Runtime cost is about US$0.005 per reply.
Mixed
Rajesh Beri, enterprise AI practitioner and author of THE DAILY BRIEF
Bedrock vs Vertex AI vs Foundry: Your Cloud Commit DecidesNewsletter analysis, THE DAILY BRIEF
The author rates Foundry the strongest of the three platforms for GPT models, with the most flexible provisioned throughput, and the weakest for Claude: no Claude provisioned throughput, no EU data zone, and Anthropic as the data processor. At list prices the platforms differ by at most about US$300 a month on a fixed Claude workload, so existing committed spend should decide.
Positive
kasey_junk, developer
DeepInfer: High-performance inference framework for DeepSeek modelsHacker News comment
The author can deploy open-weight models in Foundry with one click, already approved by the company's compliance and procurement teams. Data contracts are why the author expects cloud platforms to win enterprise model business.
Mixed or negative, 2
Negative
sebringj, developer
Runtime validation is still fucked in AI coding agentsHacker News comment
Two days of setting up AI end-to-end tests with vision calls in Foundry ran up a bill of more than US$100, which surprised the author's manager.
Negative
Julian Goldsmith (juliangoldsmith), developer
The LLM Lobotomy?Hacker News comment
Foundry still lacked tool calling for Llama models more than six months after Microsoft said it was working on it, and small models were consistently slow. Provisioned throughput would fix speed but started at about $10k a month for GPT models and about $72k for DeepSeek, against a few dollars a month of current use.
Counted for All practitioners.
Positive, 3
Positive
cdnsteve, developer
OpenCode 2.0Hacker News comment
The author's organization of more than 100 developers hosts its own models in Foundry and points Claude Code at them. Developers need no API keys because they sign in with az login.
Mixed
VanillaCache, developer
The author calls the Foundry platform really good but feels every call through it is about twice as slow.
Positive
dandinu, developer
I built a WhatsApp clone of myself; a friend agreed on a password with itHacker News comment
The author built a WhatsApp bot that answers as the author, with Claude Sonnet 4.6 through Foundry generating each reply, in about two hours of work. Runtime cost is about US$0.005 per reply.
Mixed or negative, 7
Mixed
mvATM99, developer
Kimi K2.7 Code is generally available in GitHub CopilotHacker News comment
Adding a Foundry model as a GitHub Copilot custom model failed for the author, and the path is poorly documented and unstable. VS Code's bring-your-own-key route worked, so the author now uses DeepSeek V4 Flash with Copilot.
Negative
koakuma-chan, developer
Google AntigravityHacker News comment
The author says enterprise work is forced onto cloud platforms such as Foundry, which do the same job as the vendors' own APIs with a much worse developer experience, and calls their extra prompt management and evaluation features poor.
Negative
Konnor (konnor-willison-deepsee), developer
azure-ai-projects: gpt-6-luna returns 500 on project endpoint, works on resource endpointGitHub issue, Azure/azure-sdk-for-python
A newly released model returned HTTP 500 errors on the Foundry project endpoint while working on the resource endpoint, and the failure reproduced with raw HTTP. Other models in the same project worked on both.
Negative
arielcostas, developer
Getting a Gemini API key is an exercise in frustrationHacker News comment
The author found the setup confusing: a choice between Azure OpenAI and Microsoft Foundry resources, a separate portal with an old and a new Foundry experience, Foundry and Foundry Project resource types, and a single API key unless Entra ID is set up.
Negative
Daryl Graves, developer
Microsoft's template for deploying Foundry on a private network failed both on a work subscription with an existing virtual network and on a clean personal subscription.
Negative
Oras, developer
400 reasons to not use Microsoft AzureHacker News comment
Prompt flow went down for a day with no status or information, and creating a new pay-as-you-go deployment meant switching back to the old portal to find the button.
Negative
vemonet, developer
DeepSeek R1 Is Now Available on Azure AI Foundry and GitHubHacker News comment
DeepSeek R1 on Foundry took 80 seconds to answer a question GPT-4o answered in 7, under rate limits and a 4,000-token input cap at launch. The author also found errors in the inference SDK's documentation and choppy streaming from Mistral Large.
GILD's fact checks, which don't count as sources
Checked Oct 28, 2026.Microsoft lists IDE integrations for Foundry: developers build, test, and deploy AI applications with Foundry Toolkit for Visual Studio Code. Microsoft Foundry
Checked Oct 28, 2026.Microsoft states that Foundry agents act on business systems through more than 1,400 prebuilt connections in Azure Logic Apps, including SAP, Salesforce, and Dynamics 365, or through custom tools over MCP. Microsoft Foundry
Open them in the recordEvery concern practitioners raise, most raised first.
Counted for All practitioners.
Negative
guramrit-dhillon, developer
[Azure.AI.AgentServer.Responses] response.completed is streamed before the conversation lock is released, so the next turn gets 409 conversation_lockedGitHub issue, Azure/azure-sdk-for-net
A Foundry hosted agent streams its completion event before it releases the conversation, which stays locked for a further 0.5 to 3.7 seconds. An immediate follow-up turn fails with 409 conversation_locked.
Negative
lucianteodorescu-milestone, developer
.NET: [FoundryAgent][gpt-6-luna]: [Bug]: Invalid parameter: 'text.format' of type 'json_schema' is not supported with model version gpt-6-luna-2026-09-22GitHub issue, microsoft/agent-framework
A Foundry agent that worked on one model failed with a structured output error after switching to a newer one, and the reasoning effort setting could not be removed in the Foundry agent interface.
Negative
Jino George, who published a Foundry agent to Microsoft Teams for users in his organization's Microsoft 365 tenant
Azure AI Foundry Agent published to Microsoft Teams prompts users for Foundry login instead of using Teams SSOMicrosoft Q&A thread
The author published a Foundry agent to Teams for users in his organization's tenant. On first use each user gets a sign-in card that sends them to Foundry's own sign-in service before the agent answers, where Copilot Studio agents start without one. A Microsoft moderator replied on Jun 18, 2026 that this first sign-in is expected and that no documented setting removes it.
Negative
Alexander Danson, Product Manager, Marshall Group
[Bug] 400 Error when chatting with an Agent published to Teams/Web using new foundry portalGitHub issue, microsoft-foundry/new-foundry-portal
Two agents that worked in the playground and published without error returned 400 Bad Request in Teams and web chat, with no traces to debug.
Negative
Ilya Rakhlin (irakhlin), developer
azure-ai-agents connected agent with MCP will raise for approval and eventually crashGitHub issue, Azure/azure-sdk-for-python
When a coordinator agent invoked a connected agent that used an MCP tool, the run hung in progress and then failed with a premature end of the response.
Negative
reesmonty (jacobreesmontgomery), developer
azure-ai-agents: MCP usage errors with Azure AI FoundryGitHub issue, Azure/azure-sdk-for-python
Foundry agents frequently raised errors when using MCP tools and repeatedly failed to call them. The issue drew 47 comments.
Negative
Sashank Neelamraju (sashanknjs), developer
Azure AI Foundry Agent Service Deep Research Tool ErrorGitHub issue, Azure/azure-sdk-for-python
With access to o3-deep-research granted, Microsoft's own example code for the Deep Research tool still failed with an error that the tool was not supported.
Negative
Vitor Arrais, engineer, N26
Tool call of AzureAISearchTool doesn't respect tool configuration such as filter and top_k in package azure-ai-agentsGitHub issue, Azure/azure-sdk-for-python
The agent's Azure AI Search tool ignored the filter and top_k settings configured on it.
Negative
Daryl Graves, developer
Microsoft's template for deploying Foundry on a private network failed both on a work subscription with an existing virtual network and on a clean personal subscription.
GILD's fact checks, which don't count as sources
Checked Oct 28, 2026.Foundry Agent Service connects agents to remote Model Context Protocol (MCP) server endpoints through its MCP tool, so an agent's model can call tools and data sources that other developers and organizations host. Connect agents to MCP server endpoints
Open them in the recordCounted for All practitioners.
Negative
cdnsteve, developer
DeepSeek v4.1 FlashHacker News comment
The author finds Foundry slow to add new models and missing many open-weight ones.
Negative
Baptiste GHO, CTO, Complement Learning
Baptiste GHO on X: "background mode is disabled for every model we've tried in Azure AI Foundry"Post on X
Background mode was disabled for every model the company tried in Foundry, which blocked a new course generation feature in production.
Negative
Julian Goldsmith (juliangoldsmith), developer
The LLM Lobotomy?Hacker News comment
Foundry still lacked tool calling for Llama models more than six months after Microsoft said it was working on it, and small models were consistently slow. Provisioned throughput would fix speed but started at about $10k a month for GPT models and about $72k for DeepSeek, against a few dollars a month of current use.
Mixed
Rajesh Beri, enterprise AI practitioner and author of THE DAILY BRIEF
Bedrock vs Vertex AI vs Foundry: Your Cloud Commit DecidesNewsletter analysis, THE DAILY BRIEF
The author rates Foundry the strongest of the three platforms for GPT models, with the most flexible provisioned throughput, and the weakest for Claude: no Claude provisioned throughput, no EU data zone, and Anthropic as the data processor. At list prices the platforms differ by at most about US$300 a month on a fixed Claude workload, so existing committed spend should decide.
Negative
lucianteodorescu-milestone, developer
.NET: [FoundryAgent][gpt-6-luna]: [Bug]: Invalid parameter: 'text.format' of type 'json_schema' is not supported with model version gpt-6-luna-2026-09-22GitHub issue, microsoft/agent-framework
A Foundry agent that worked on one model failed with a structured output error after switching to a newer one, and the reasoning effort setting could not be removed in the Foundry agent interface.
Mixed
mvATM99, developer
Kimi K2.7 Code is generally available in GitHub CopilotHacker News comment
Adding a Foundry model as a GitHub Copilot custom model failed for the author, and the path is poorly documented and unstable. VS Code's bring-your-own-key route worked, so the author now uses DeepSeek V4 Flash with Copilot.
Negative
Sashank Neelamraju (sashanknjs), developer
Azure AI Foundry Agent Service Deep Research Tool ErrorGitHub issue, Azure/azure-sdk-for-python
With access to o3-deep-research granted, Microsoft's own example code for the Deep Research tool still failed with an error that the tool was not supported.
Negative
vemonet, developer
DeepSeek R1 Is Now Available on Azure AI Foundry and GitHubHacker News comment
DeepSeek R1 on Foundry took 80 seconds to answer a question GPT-4o answered in 7, under rate limits and a 4,000-token input cap at launch. The author also found errors in the inference SDK's documentation and choppy streaming from Mistral Large.
Counted for All practitioners.
Mixed
VanillaCache, developer
The author calls the Foundry platform really good but feels every call through it is about twice as slow.
Negative
Konnor (konnor-willison-deepsee), developer
azure-ai-projects: gpt-6-luna returns 500 on project endpoint, works on resource endpointGitHub issue, Azure/azure-sdk-for-python
A newly released model returned HTTP 500 errors on the Foundry project endpoint while working on the resource endpoint, and the failure reproduced with raw HTTP. Other models in the same project worked on both.
Mixed
mrsmrtss, developer
Decisions that eroded trust in Azure - by a former Azure Core engineerHacker News comment
The author's team has run critical services on Azure for ten years without a major incident, but calls the AI Foundry side broken at the moment, as are other vendors' AI services.
Negative
hobofan, developer
Getting a Gemini API key is an exercise in frustrationHacker News comment
The author sees Foundry serving return HTTP 500 errors for stretches of time and advises production teams to pick models available from several clouds and fail over between them.
Negative
vemonet, developer
DeepSeek R1 Is Now Available on Azure AI Foundry and GitHubHacker News comment
DeepSeek R1 on Foundry took 80 seconds to answer a question GPT-4o answered in 7, under rate limits and a 4,000-token input cap at launch. The author also found errors in the inference SDK's documentation and choppy streaming from Mistral Large.
Negative
Julian Goldsmith (juliangoldsmith), developer
The LLM Lobotomy?Hacker News comment
Foundry still lacked tool calling for Llama models more than six months after Microsoft said it was working on it, and small models were consistently slow. Provisioned throughput would fix speed but started at about $10k a month for GPT models and about $72k for DeepSeek, against a few dollars a month of current use.
Negative
Oras, developer
400 reasons to not use Microsoft AzureHacker News comment
Prompt flow went down for a day with no status or information, and creating a new pay-as-you-go deployment meant switching back to the old portal to find the button.
Counted for All practitioners.
Raised as a concern, 6
Mixed
AiswaryaVenkit1, developer
The author says Foundry can feel at first like a long list of models, agents, tools, evaluations, guardrails, and governance features.
Negative
arielcostas, developer
Getting a Gemini API key is an exercise in frustrationHacker News comment
The author found the setup confusing: a choice between Azure OpenAI and Microsoft Foundry resources, a separate portal with an old and a new Foundry experience, Foundry and Foundry Project resource types, and a single API key unless Entra ID is set up.
Negative
Daryl Graves, developer
Microsoft's template for deploying Foundry on a private network failed both on a work subscription with an existing virtual network and on a clean personal subscription.
Negative
Oras, developer
400 reasons to not use Microsoft AzureHacker News comment
Prompt flow went down for a day with no status or information, and creating a new pay-as-you-go deployment meant switching back to the old portal to find the button.
Negative
Alexander Danson, Product Manager, Marshall Group
[Bug] 400 Error when chatting with an Agent published to Teams/Web using new foundry portalGitHub issue, microsoft-foundry/new-foundry-portal
Two agents that worked in the playground and published without error returned 400 Bad Request in Teams and web chat, with no traces to debug.
Mixed
zubiaur, developer
Microsoft has a problem: lack of demand for its AI productsHacker News comment
The author calls Azure's AI services decent and access to OpenAI models through Foundry a real differentiator, but says rate limits are a problem. The author worries the Foundry rebranding signals marketing taking over the services.
Discussed without concern, 1
Mixed
chris_money202, developer
GPT 5.6 Sol Still Fails on Basic Code ReviewHacker News comment
Checking the Foundry catalog, the author found that the separate reasoning variants of a model seemed to be gone, leaving what may be one model.
GILD's fact checks, which don't count as sources
Checked Oct 28, 2026.Microsoft states the platform moved from Azure AI Studio to Azure AI Foundry to Microsoft Foundry while the Azure resource type stayed Microsoft.CognitiveServices/accounts. The Assistants API retired on August 26, 2026, and the Responses API that agents now use is not available in every Azure region. Migrate from the Foundry (classic) portal
Open them in the recordCounted for All practitioners.
Negative
sebringj, developer
Runtime validation is still fucked in AI coding agentsHacker News comment
Two days of setting up AI end-to-end tests with vision calls in Foundry ran up a bill of more than US$100, which surprised the author's manager.
Mixed
zubiaur, developer
Microsoft has a problem: lack of demand for its AI productsHacker News comment
The author calls Azure's AI services decent and access to OpenAI models through Foundry a real differentiator, but says rate limits are a problem. The author worries the Foundry rebranding signals marketing taking over the services.
Negative
Julian Goldsmith (juliangoldsmith), developer
The LLM Lobotomy?Hacker News comment
Foundry still lacked tool calling for Llama models more than six months after Microsoft said it was working on it, and small models were consistently slow. Provisioned throughput would fix speed but started at about $10k a month for GPT models and about $72k for DeepSeek, against a few dollars a month of current use.
Mixed
Rajesh Beri, enterprise AI practitioner and author of THE DAILY BRIEF
Bedrock vs Vertex AI vs Foundry: Your Cloud Commit DecidesNewsletter analysis, THE DAILY BRIEF
The author rates Foundry the strongest of the three platforms for GPT models, with the most flexible provisioned throughput, and the weakest for Claude: no Claude provisioned throughput, no EU data zone, and Anthropic as the data processor. At list prices the platforms differ by at most about US$300 a month on a fixed Claude workload, so existing committed spend should decide.
Negative
vemonet, developer
DeepSeek R1 Is Now Available on Azure AI Foundry and GitHubHacker News comment
DeepSeek R1 on Foundry took 80 seconds to answer a question GPT-4o answered in 7, under rate limits and a 4,000-token input cap at launch. The author also found errors in the inference SDK's documentation and choppy streaming from Mistral Large.
GILD's fact checks, which don't count as sources
Checked Oct 28, 2026.Foundry Models quotas apply per Azure subscription and region, and increases go through a request form evaluated one by one. Spillover sends overage from a full provisioned deployment to a pay-as-you-go deployment of the same model and version. Microsoft Foundry Models quotas and limits
Open them in the recordCounted for All practitioners.
Raised as a concern, 3
Mixed
Rajesh Beri, enterprise AI practitioner and author of THE DAILY BRIEF
Bedrock vs Vertex AI vs Foundry: Your Cloud Commit DecidesNewsletter analysis, THE DAILY BRIEF
The author rates Foundry the strongest of the three platforms for GPT models, with the most flexible provisioned throughput, and the weakest for Claude: no Claude provisioned throughput, no EU data zone, and Anthropic as the data processor. At list prices the platforms differ by at most about US$300 a month on a fixed Claude workload, so existing committed spend should decide.
No opinion stated
browningstreet, developer at a large enterprise
Why isn't the industry freaking out about DeepSeek 4.1 Flash?Hacker News comment
The author's large enterprise bans Chinese models outright and does not yet treat a Chinese model hosted in Foundry any differently.
Negative
jiggawatts, consultant
GPT-6 Astra on OpenRouterHacker News comment
The author says zero data retention is the main reason customers pick Foundry over OpenAI's own API, yet the model list makes it hard to see which models carry it. There is no simple way to block models without it, so each customer vets and approves models one by one.
Discussed without concern, 3
No opinion stated
HanClinto, developer
DeepSeek v4.1 FlashHacker News comment
The author explains that changing or removing Foundry's safety guardrails and content filters is a Limited Access feature that requires a request form.
No opinion stated
rombert, developer
Kimi K2.7 Code is generally available in GitHub CopilotHacker News comment
Quoting GitHub's documentation, the author points out that these models run on US-based Foundry infrastructure and that prompts and responses are not sent to the original model developers.
Positive
consumer451, developer building products for clients
Cybersecurity researchers aren't happy about the guardrails on Anthropic's FableHacker News comment
The author deploys Claude models on Azure because clients already trust their Microsoft contracts, and says the Anthropic models make the products much better. The author asks how data terms for newer frontier models affect that trust.
GILD's fact checks, which don't count as sources
Checked Oct 28, 2026.For Foundry Models sold by Azure, prompts, completions, embeddings, and training data are not used to train generative AI foundation models without the customer's permission, and Data Zone deployments keep processing inside the data zone. Data, privacy, and security for Foundry Models sold by Azure in Microsoft Foundry
Open them in the recordStatements this group's sources agree on, then the ones close to agreement.
Teams use Foundry when their company already trusts its Microsoft agreement, needs zero data retention, or must route models through approved compliance and procurement.
10 supporting sourcesClose to agreement: each needs 9 supporting sources
Agent features that work in the playground or in Microsoft's own examples fail in teams' own setups, so teams test each tool, model switch, and publishing path before production.
7 supporting sourcesQuestions the sources split on, with the count on each side.
Evidence remains mixed
Broad enough, 6 sources
Some find nearly every model they need, from OpenAI to DeepSeek, approved and ready to deploy.
Gaps and delays, 4 sources
Others find new and open-weight models arrive late, or arrive without features such as tool calling or reserved capacity.
The three groups, then every practitioner by the context they work in, and how emphasis changes with scale.
Three groups drawn from the same sources. Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. Signals open on All practitioners. Teams and leaders have 8 accounts; GILD prints their sentiment from 20.
10 sources. These accounts come from teams and companies building on Foundry and focus on data terms and on running agents in production.
35 sources. These accounts focus on the model catalog and on Foundry Agent Service, from developers building on Foundry on their own.
45 sources. This gives the broadest read, but combines solo projects with production teams that can reach different conclusions.
45 sources from practitioners, by the context they work in. Filled rows count for this group.
Emphasizes Catalog breadth and Agent Service bugs
Emphasizes Reliability and unexpected bills
Emphasizes Agent tools and publishing in production
Emphasizes Platform choice by committed spend, and missing model features
Emphasizes Keyless access and model policy at scale
What changes with scale. Individual builders focus on which models the catalog carries and on Agent Service bugs. Team and company accounts focus on data terms under existing Microsoft agreements and on features that block production use.
Where this group's sources were published.
Counted for All practitioners, from 45 sources. Change group
59 sources for this group, 14 of them documentation
Older evidence stays when it still applies. Accounts tied to the classic portal, retired agent APIs, or models that have since changed remain in the history with a Needs a recheck label.
See sources by originWho each source is, how it knows, whether it still applies and what it counts toward.
Newest first, on its own page. Every count above opens its sources there.
Read all 59 sourcesAbout this read
A running review of what teams building AI features on Microsoft Foundry report, from individual builders to company-wide platforms. Each finding links to its sources, and each conflict shows both sides.
Counts cover only the sources GILD reviewed.
Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. GILD files each source in a group from how the account describes its author. A role or employer marked self-described is what the source states about themselves. Documentation checks facts for every group and never counts toward sentiment.
A source's date and whether it still applies are separate: a recent post can describe an old version, and an old one can still hold.
Flagged for a recheck, of the sources that take a side: teams and leaders 1 of 8, individual builders 7 of 28, all practitioners 8 of 36.
Each row opens its sources for the group you are reading.
| Finding | Teams and technical leaders | Individual builders | All practitioners | Sources |
|---|---|---|---|---|
| Current readCurrent sentiment | 1 positive, 2 mixed, 5 negative | 8 positive, 5 mixed, 15 negative | 9 positive, 7 mixed, 20 negative | Sources |
| PraiseModel Catalog & Choice | 1 of 2 positive (under the minimum) | 10 of 14 positive | 11 of 16 positive | Sources |
| PraiseContracts, Data Terms & Identity | 2 of 2 positive (under the minimum) | 6 of 8 positive | 8 of 10 positive | Sources |
| PraiseCost | 1 of 2 positive (under the minimum) | 2 of 3 positive (under the minimum) | 3 of 5 positive | Sources |
| PraiseDeveloper Experience | 1 of 1 positive (under the minimum) | 2 of 9 positive | 3 of 10 positive | Sources |
| ConcernAgent Service & Tool Failures | 3 of 3 raise it (under the minimum) | 6 of 6 raise it | 9 of 9 raise it | Sources |
| ConcernReliability & Latency | 1 of 1 raise it (under the minimum) | 6 of 6 raise it | 7 of 7 raise it | Sources |
| ConcernPortal, Naming & Setup | 1 of 1 raise it (under the minimum) | 5 of 6 raise it | 6 of 7 raise it | Sources |
| ConcernMissing Models & Features | 2 of 2 raise it (under the minimum) | 6 of 6 raise it | 8 of 8 raise it | Sources |
| ConcernData Terms & Model Governance | 2 of 2 raise it (under the minimum) | 1 of 4 raise it (under the minimum) | 3 of 6 raise it | Sources |
| ConcernQuotas, Capacity & Cost | 2 of 2 raise it (under the minimum) | 3 of 3 raise it (under the minimum) | 5 of 5 raise it | Sources |
| ConsensusChosen for the contract and data terms | 2 supporting (under the minimum) | 8 supporting | 10 supporting | Sources |
| ConsensusProve agents on your own setup | 2 supporting (under the minimum) | 5 supporting (below threshold) | 7 supporting (below threshold) | Sources |
| DisagreementDoes Foundry's catalog have the models a team needs? | Broad enough 0, Gaps and delays 1 (under the minimum) | Broad enough 6, Gaps and delays 3 | Broad enough 6, Gaps and delays 4 | Sources |
| Current readEvidence mix | 24 sources | 49 sources | 59 sources | Sources |
GILD's checks of claims against official documentation, kept apart from the source counts.
