GPT-5.4 Lands in Microsoft Foundry: What It Actually Means for Your .NET Agent Pipelines

TL;DR

OpenAI’s GPT-5.4 is now generally available in Microsoft Foundry, bringing a 1-million-token context window, a redesigned Tool Search system, improved instruction adherence, and measurably fewer hallucinations — all accessible through the Azure.AI.Projects SDK in C#. If you’re shipping agentic workflows on Azure, this is the model upgrade worth planning for right now.

What Just Shipped

Today, GPT-5.4 is generally available in Microsoft Foundry — a model designed to help organizations move from planning work to reliably completing it in production environments. That’s not marketing boilerplate. It’s a meaningful shift in emphasis: previous GPT generations were optimized for correctness on benchmarks; GPT-5.4 is explicitly optimized for finishing the job in multi-step, multi-tool workflows. In addition to the standard version, GPT-5.4 is also available as a reasoning model (GPT-5.4 Thinking) or optimized for high performance (GPT-5.4 Pro). The API version supports context windows as large as 1 million tokens — by far the largest context window available from OpenAI. And yes, you read that right: one million tokens. Drop your entire codebase in there. Go ahead. We’ll wait.

The Numbers Engineers Actually Care About

Hallucination Reduction

GPT-5.4 is 33% less likely to make errors in individual claims when compared to GPT-5.2, and overall responses are 18% less likely to contain errors. For anyone building RAG pipelines over enterprise data — legal documents, financial reports, internal wikis — this is the kind of stat that translates directly into fewer angry Slack messages from the compliance team.

Latency and Token Efficiency

OpenAI also emphasized improved token efficiency, saying GPT-5.4 was able to solve the same problems with significantly fewer tokens than its predecessor. Fewer tokens in means lower cost and faster time-to-first-token. Latency-improved performance for responsive, real-time workflows is explicitly called out as one of GPT-5.4’s production-grade advancements.

Agentic Execution

GPT-5.4 combines stronger reasoning with built-in computer use capabilities to support automation scenarios, and dependable execution across tools, files, and multi-step workflows at scale. The model’s more dependable tool invocation reduces prompt tuning and human oversight — meaning your agents should require fewer “please do exactly what I said” system-prompt gymnastics.

Tool Search — A New API Primitive

As part of the launch, OpenAI has reworked how the API version of GPT-5.4 manages tool calling, introducing a new system called Tool Search. This is worth paying close attention to: instead of your application maintaining a static list of tools and hoping the model picks the right one, Tool Search lets the model discover the most relevant tool for a given task from a larger registered set. Less prompt engineering, more reliable tool selection.

GitHub Copilot Gets It Too

If you’re using GitHub Copilot in Visual Studio or VS Code, GPT-5.4, OpenAI’s latest agentic coding model, is now rolling out in GitHub Copilot. In early testing of real-world, agentic, and software development capabilities, GPT-5.4 consistently hits new rates of success. It also shows enhanced logical reasoning and task execution for intricate, multi-step, tool-dependent processes. You can select the model in the model picker in Visual Studio Code v1.104.1 and later (all modes: chat, ask, edit, agent) and Visual Studio version 17.14.19 and later (agent, ask modes).

Note for Enterprise/Business admins: Copilot Enterprise and Copilot Business plan administrators must enable the GPT-5.4 policy in Copilot settings before it shows up for your developers. Check your org policy before wondering why the new model isn’t appearing.

Calling GPT-5.4 from .NET

The Foundry SDK has been consolidating fast. The azure-ai-projects v2 line is the new canonical SDK for everything Foundry: agents (now built on the OpenAI Responses protocol), evaluations, memory stores, and model inference. Here’s how you’d call GPT-5.4 in C# using the Foundry SDK today (swap gpt-5.2 for gpt-5.4 once your registration is approved — the pattern is identical):

using Azure.AI.Projects.OpenAI;
using Azure.Identity;
using OpenAI.Responses;

string endpoint = "https://<resource-name>.services.ai.azure.com/api/projects/<project-name>";

AIProjectClient projectClient = new(
    endpoint: new Uri(endpoint),
    tokenProvider: new DefaultAzureCredential());

// Swap model name to "gpt-5.4" or "gpt-5.4-pro" when available in your subscription
OpenAIResponseClient responseClient =
    projectClient.OpenAI.GetProjectResponsesClientForModel("gpt-5.4");

OpenAIResponse response = responseClient.CreateResponse(
    "Summarize the key risk clauses in the attached contract.");

Console.WriteLine(response.GetOutputText());

Authentication uses DefaultAzureCredential — Managed Identity in production, your local az login in dev. No API keys floating around in appsettings.json. Your security team will appreciate this.

Access and Availability: The Catch

Registration is required for access to gpt-5.4 and gpt-5.4-pro. Customers who previously applied and received access to a limited access model don’t need to reapply, as their approved subscriptions will automatically be granted access upon model release. If you’re starting fresh, head to the Microsoft Foundry Model Catalog and apply. Approval timelines vary, so don’t wait until the sprint you actually need it.

Also worth flagging for anyone still on older models: the gpt-4o-mini model is set to retire from Azure Foundry. According to available information, the retirement date is March 31, 2026. You will need to switch to a different model before this date to avoid any disruption in your production environment. That’s two weeks from today. If you have production workloads on gpt-4o-mini in Azure Foundry, stop reading this and go check your deployments. Seriously.

Practical Takeaways

What	Action
GPT-5.4 GA in Foundry	Request access now via the Foundry Model Catalog
1M token context	Revisit chunking strategies — you may not need them
Tool Search API	Redesign large tool registries to leverage dynamic discovery
GitHub Copilot	Enable GPT-5.4 policy in your org admin settings
`gpt-4o-mini` retirement	Migrate Azure Foundry deployments before March 31, 2026
`azure-ai-projects` v2	Adopt the unified SDK; `azure-ai-agents` as a standalone package is gone

One More Thing: Model Lifecycle Is Its Own Project

OpenAI (ChatGPT) and Azure AI Foundry operate on independent lifecycle policies. When OpenAI announces that a model is retired from ChatGPT or OpenAI-hosted APIs, that retirement does not automatically apply to Azure AI Foundry. Always consult the Foundry portal’s Model Catalog → Retirement date column for Azure-authoritative timelines — not OpenAI’s blog posts.

The model churn in 2026 has been relentless (GPT-5, 5.1, 5.2, 5.3, and now 5.4 — all within months). Building your integration against a pinned deployment name rather than a floating alias, and treating model upgrades as first-class engineering work in your backlog, is no longer optional hygiene. It’s table stakes.