Microsoft Foundry−26 Negative

Operator Insights

What are teams learning from building on Microsoft Foundry and running it in production?

4 accounts: 2 failures, incidents or rollbacks and 2 adopted it

Failures, incidents and rollbacks2

Chris BakerFurther's About page lists Chris Baker as COO. The post's author says he is a co-founder and signs it With Further Inc.Further Inc.
A 13-person startup in the Microsoft for Startups Founders Hub program, migrating its AI infrastructure.
Microsoft Tech Community, Startups at Microsoft, published Apr 13, 2026Read the post (opens Microsoft Tech Community, Startups at Microsoft in a new tab)
Observation
Further deployed Claude Opus 4.6 and Sonnet 4.6 through Azure AI Foundry assuming its sponsorship credits covered them, and was invoiced $17,493.01 after Foundry failed to warn that these models are excluded from the credits and left their cost estimate blank.
Takeaway
Before you deploy Claude or another partner model in Foundry on sponsorship credits, confirm the credits cover it and watch the invoices, because this team found no exclusion warning and no cost estimate for those models.
What happened
In February 2026 Further deployed Claude Opus 4.6 and Sonnet 4.6 through Azure AI Foundry as part of a migration of its AI infrastructure, assuming its Azure sponsorship credits covered them. The post says Foundry showed no warning or label that these models are excluded from sponsorship credits, and that the monitoring dashboard's Estimated Cost section was blank for them. Invoices of $1,078.07 for February and $16,414.94 for March followed, $17,493.01 in all; the post says Further could not pay the second, and that Azure Support, Anthropic and the Microsoft for Startups team each referred it elsewhere.
- Public organization write-up (post signed With Further Inc.)
- Posted Apr 13, 2026
Deployment details
A 13-person startup in the Microsoft for Startups Founders Hub program, migrating its AI infrastructure.
- Public organization write-up (post signed With Further Inc.)
- Posted Apr 13, 2026
Team
Further Inc., a 13-person startup
Workload
A migration of the company's AI infrastructure
Product version
Claude models deployed through Azure AI Foundry, expected to run on Azure sponsorship credits
Model
Claude Opus 4.6 and Sonnet 4.6
Measurement
Invoices of $1,078.07 (February) and $16,414.94 (March), $17,493.01 in all

ONTHEBIASJul 10, 2026Microsoft Q&AONTHEBIAS could not create any new Azure OpenAI model deployment: every gpt-5.1 attempt failed with error 715-123420 although Microsoft had approved the quota, and its founder says the restriction blocks the company's production rollout.
Jack WieseFounderONTHEBIAS
ONTHEBIAS's design-to-manufacturing platform, for which the company needs Azure OpenAI as its production AI provider.
Microsoft Q&A, published Jul 10, 2026Read the question (opens Microsoft Q&A in a new tab)
Observation
ONTHEBIAS could not create any new Azure OpenAI model deployment: every gpt-5.1 attempt failed with error 715-123420 although Microsoft had approved the quota, and its founder says the restriction blocks the company's production rollout.
Takeaway
Create the Foundry deployments a launch depends on well before the date, because an account restriction can block new deployments while quota is approved and existing deployments keep serving.
What happened
Every attempt to create a new gpt-5.1 Global Standard deployment, on two resources in East US and East US 2, failed with error 715-123420, which the post calls a false-positive fraud-protection restriction. Microsoft had approved the quota, and an existing gpt-5-mini deployment kept working. On Aug 5, 2026 a moderator marked Microsoft External Staff commented that they had reached out by private message; the thread shows no resolution.
Public organization write-up (founder's Microsoft Q&A question)
Error code, regions and quota state stated
Asked Jul 10, 2026
Deployment details
ONTHEBIAS's design-to-manufacturing platform, for which the company needs Azure OpenAI as its production AI provider.
Public organization write-up (founder's Microsoft Q&A question)
Error code, regions and quota state stated
Asked Jul 10, 2026
Team
ONTHEBIAS
Workload
Production AI provider for a design-to-manufacturing platform
Product version
New Global Standard deployments on two Azure OpenAI resources, East US and East US 2
Model
gpt-5.1 blocked; an existing gpt-5-mini deployment kept working

Adopted it2

Swarup S. SwaminathanAssociate Professor of Clinical Ophthalmology; the study's senior and corresponding authorBascom Palmer Eye Institute, University of Miami Miller School of Medicine
A University of Miami research team reading medications from glaucoma clinic notes, inside the university's existing Microsoft Azure enterprise environment.
Ophthalmology Science, on PubMed Central, published Aug 17, 2026Read the paper (opens Ophthalmology Science, on PubMed Central in a new tab)
Observation
Bascom Palmer Eye Institute researchers used Microsoft Azure AI Foundry to run five models on 992 glaucoma clinic notes under a Microsoft business associate agreement, and failures ranged from 0% to 0.9%.
Takeaway
A hospital can run models from several makers on patient data through one Foundry environment under its Microsoft business associate agreement. Track token use and cost from the first run, which this study did not.
What happened
The researchers drew 1,250 glaucoma clinic notes at random, one per patient, from 451,283 eligible notes, excluded 8 supervising-physician attestations to leave 1,242, and used 992 as the held-out test set for every model performance analysis. Claude Opus 4.6, DeepSeek-V3.2, GPT-5.2, Grok 4.1 and Qwen3.6-35B-A3B were accessed through Microsoft Azure AI Foundry inside HIPAA-compliant containers, under a business associate agreement between Microsoft and the University of Miami. On the held-out set, F1 for current topical medications ranged from 0.968 (Grok) to 0.985 (Claude), failures ranged from 0 to 9 (0% to 0.9%) across models and tasks, all from API timeouts or standardization errors, and token usage and costs were not tracked.
- Public engineering write-up (peer-reviewed paper in Ophthalmology Science)
- Measured results on stated settings
- Published Aug 17, 2026
Deployment details
A University of Miami research team reading medications from glaucoma clinic notes, inside the university's existing Microsoft Azure enterprise environment.
- Public engineering write-up (peer-reviewed paper in Ophthalmology Science)
- Measured results on stated settings
- Published Aug 17, 2026
Team
Bascom Palmer Eye Institute researchers
Workload
Reading current topical and oral medications from glaucoma clinic notes; 992 notes in the held-out test set
Product version
Microsoft Azure AI Foundry, in HIPAA-compliant containers within the university's existing Azure enterprise environment
Model
Claude Opus 4.6, DeepSeek-V3.2, GPT-5.2, Grok 4.1 and Qwen3.6-35B-A3B
Measurement
F1 for current topical medications from 0.968 (Grok) to 0.985 (Claude); failures from 0 to 9 (0% to 0.9%) across models and tasks; token usage and costs not tracked

Canadian Food Inspection Agency AI LabMar 24, 2026GitHub, ai-cfia/fertiscanThe CFIA AI Lab reported that Azure AI Foundry was retiring the GPT-4o deployment FertiScan uses, and asked for an upgrade to a supported model of version 5.1 or later.
Kotchikpa Guy-Landry AllagbeFertiScan datastore package co-authorCanadian Food Inspection Agency AI Lab
Public sector. FertiScan, a CFIA AI Lab prototype under active development that inspectors are trying out on real fertilizer labels.
GitHub, ai-cfia/fertiscan, published Mar 24, 2026Read the issue (opens GitHub, ai-cfia/fertiscan in a new tab)
Observation
The CFIA AI Lab reported that Azure AI Foundry was retiring the GPT-4o deployment FertiScan uses, and asked for an upgrade to a supported model of version 5.1 or later.
Takeaway
Budget for model migration and validation when you build on a Foundry model deployment, since a retirement can force an upgrade.
What happened
On Mar 24, 2026 the lab opened an issue in its FertiScan repository saying Azure AI Foundry was retiring its current GPT-4o deployment and that it needed a supported model of version 5.1 or later. The issue was still open on Oct 30, 2026. Its outcome and acceptance criteria describe the upgrade the lab wants, and an Apr 13, 2026 comment asking whether the switch happened automatically has no reply.
Public engineering write-up (issue in the CFIA AI Lab's own GitHub repository)
Opened Mar 24, 2026
Deployment details
Public sector. FertiScan, a CFIA AI Lab prototype under active development that inspectors are trying out on real fertilizer labels.
Public engineering write-up (issue in the CFIA AI Lab's own GitHub repository)
Opened Mar 24, 2026
Team
CFIA AI Lab, FertiScan project
Workload
FertiScan, a prototype inspectors are trying out on real fertilizer labels
Product version
A GPT-4o deployment in Azure AI Foundry
Model
GPT-4o

Failures, incidents and rollbacks2

Further deployed Claude Opus 4.6 and Sonnet 4.6 through Azure AI Foundry assuming its sponsorship credits covered them, and was invoiced $17,493.01 after Foundry failed to warn that these models are excluded from the credits and left their cost estimate blank.

ONTHEBIAS could not create any new Azure OpenAI model deployment: every gpt-5.1 attempt failed with error 715-123420 although Microsoft had approved the quota, and its founder says the restriction blocks the company's production rollout.

Adopted it2

Bascom Palmer Eye Institute researchers used Microsoft Azure AI Foundry to run five models on 992 glaucoma clinic notes under a Microsoft business associate agreement, and failures ranged from 0% to 0.9%.

The CFIA AI Lab reported that Azure AI Foundry was retiring the GPT-4o deployment FertiScan uses, and asked for an upgrade to a supported model of version 5.1 or later.