Microsoft has officially integrated OpenAI’s newly launched GPT-5 into Copilot, just a day after it brought the company’s open-source GPT models to its local AI services. The rollout marks one of the fastest adoptions of a major AI model into Microsoft’s flagship assistant platform.
As of Thursday, GPT-5 is live at copilot.microsoft.com, where users can manually enable it via a drop-down menu. The model is labeled as “Thinks deeply or quickly based upon the task,” a hint at its new “router” functionality, which OpenAI says can automatically direct queries to specialized subsystems depending on complexity. The same technology is already in ChatGPT, where it powers both broad and domain-specific improvements in reasoning, creative writing, coding, and even health-related tasks.
For now, the Windows Copilot app appears to still rely on GPT-4 — at least in tests run by reporters — but Microsoft suggests GPT-5 will eventually power the desktop experience as well. In addition, the company confirmed that GPT-5 will soon arrive in Microsoft 365 Copilot, GitHub Copilot, Visual Studio Code, and Azure AI Foundry. On Azure AI Foundry, GPT-5 is already available starting today, while GitHub Copilot’s upgrade will roll out later.
OpenAI claims GPT-5 can, for example, design and implement a complete website from a single prompt, reflecting a step-change in generative coding capabilities. It’s also meant to respond more flexibly, switching between rapid responses and deep reasoning depending on user needs.
Early, informal testing shows some differences between Copilot’s GPT-5 and ChatGPT’s GPT-5 in detail depth. A whimsical test prompt — “How much wood could a woodchuck chuck if a woodchuck could chuck wood?” — returned the same answer on both platforms (700 pounds), but ChatGPT provided a more elaborate breakdown.
This update follows Microsoft’s Wednesday announcement of two open-source models, gpt-oss-120B and gpt-oss-20B, on Azure AI Foundry, as well as GPU-optimized versions of the 20B model for Windows via Foundry Local. Those local models require GPUs with at least 16GB of VRAM, though advanced integrated GPU configurations — like the Ryzen AI Max+ 395 in the Framework Desktop, which can allocate up to 96GB of VRAM — have shown strong results, even outperforming a more complex Meta Llama Scout 109B model in some benchmarks.
Microsoft says GPT-5’s safety profile is stronger than any previous OpenAI model, showing resilience against harmful use cases such as malware creation, fraud automation, and other malicious exploits. If true, that could address one of the biggest ongoing concerns with rapid AI advancement: keeping capability growth matched with risk controls.
