Microsoft announced a standalone GitHub Copilot desktop app yesterday at Build 2026, alongside MAI-Code-1-Flash, a new language model purpose-built for coding assistance.

The desktop app is what GitHub is calling an “agent-native” experience. That means Copilot can now run outside your IDE and interact with your entire desktop environment, not just the text editor. The details are thin so far, but the implication is that the agent can access files, terminals, browsers, and other desktop apps without being confined to a VS Code sidebar.

Microsoft’s new code model

The more interesting part is what’s running under the hood. MAI-Code-1-Flash is a 137-billion-parameter model with 5 billion active parameters, using a mixture-of-experts architecture. Microsoft says it’s rolling out to GitHub Copilot individual users in VS Code now.

Simon Willison notes that 137B total and 5B active is remarkably small for a frontier model. For context, most recent flagship models are well into the trillions of parameters. The efficiency play here is obvious: smaller active parameter count means faster inference and lower costs, which matters when you’re running completions on every keystroke.

Microsoft also announced MAI-Thinking-1, a 1-trillion-parameter reasoning model with 35B active, but that’s limited to “select early partners” for now. No public access yet.

OpenAI counters with job-specific Codex plugins

On the same day, OpenAI released six new plugins for Codex aimed at white-collar work: data analytics, creative production, sales, product design, equity investing, and investment banking. Each plugin bundles integrations, instructions, and context specific to that job function.

This is OpenAI’s play to move Codex beyond general-purpose chat into verticalized workflows. Instead of asking Codex how to analyze a dataset, the data analytics plugin knows which tools you use, what formats you work with, and what “analyze this” actually means in your context.

The plugins are available now inside the Codex app. No word on pricing changes or whether these will eventually become separate SKUs.

Who should care

If you’re a GitHub Copilot user, the desktop app is worth watching. The current VS Code integration is good but limited. If the agent can actually operate across your whole system, that’s a meaningful upgrade. The catch is we don’t know yet what “agent-native” actually means in practice or what permissions model they’re using.

If you’re building coding tools or LLM-powered dev workflows, MAI-Code-1-Flash’s parameter efficiency is interesting. Microsoft is betting that MoE architectures with small active parameter counts can match larger models on domain-specific tasks. If they’re right, it changes the economics of AI code assistance.

If you use Codex for work that’s not software development, the new plugins might be immediately useful. If you don’t, they’re easy to ignore.

What’s not here

Microsoft didn’t share benchmarks, pricing details, or access timelines for most of this. The desktop app announcement doesn’t include a release date beyond “coming soon.” MAI-Thinking-1 is partner-only with no public roadmap. The Codex plugins shipped, but we don’t know if they actually work well or if they’re just GPT-4 with a system prompt and some API keys.

Still, the direction is clear. Both companies are moving from general-purpose chat interfaces to specialized agents that understand specific tools, workflows, and contexts. Whether they’re good enough to change how you work is the question we’ll be able to answer in a few weeks once people actually use them.