Microsoft has quietly flipped a switch inside Visual Studio Code that changes which AI model answers your coding questions. For paid GitHub Copilot users who enable the new "auto model selection" feature, Anthropic’s Claude Sonnet 4 is now the primary backend—not OpenAI’s GPT. The change, rolled out in preview, also comes with a 10% discount on premium requests when you let the system pick the model for you.

It’s a significant shift in strategy for Microsoft, which built Copilot on OpenAI’s technology. But internal tests convinced the company that Claude is simply better at many coding tasks. The move doesn’t lock out GPT models—you can still choose them manually—but for most paid users who leave the setting on "Auto," Claude is the new default brain.

How Auto Model Selection Actually Works

Behind the scenes, the auto model selector inspects each request you type into Copilot Chat. It weighs factors like the complexity of your prompt, current server load, and your available quota, then routes the entire chat session to what it considers the best model. For paid users, that routing heavily favors Claude Sonnet 4, though GPT-5, GPT-5 mini, and other models remain in the mix.

You don’t need to install anything extra—the feature is built directly into VS Code’s Copilot extension. If you’re subscribed to a Copilot Pro, Pro+, Business, or Enterprise plan, the Auto setting will start sending your requests to Anthropic’s servers through a partnership between Microsoft and Anthropic. Hovering over any response in the chat window reveals which model actually generated the answer, so you can see when Claude is at work.

Microsoft loosely calls this a "model orchestrator." The 10% premium-request discount is meant to encourage adoption: when you let the system choose, each request uses slightly fewer tokens from your monthly allowance. If you run out of premium requests, the system gracefully falls back to a free model so you’re never left without any AI assistance.

A second, more technical path exists alongside this. Anthropic offers its own Claude Code CLI and a VS Code extension that pairs directly with Claude models, bypassing Copilot entirely. With it, you can open a terminal inside VS Code, run claude, and get agentic capabilities: multi-file edits proposed as diffs, context from your open files and diagnostics, and the ability to execute commands and run tests. This route requires your own Anthropic API key and an understanding of the command line, but it gives power users and teams far more control.

What This Shift Means for Your Daily Coding

For the typical developer who already pays for Copilot, the most immediate difference is simply better answers. Claude 4’s expanded context window—capable of reasoning over tens of thousands of tokens—means you can ask it to analyze an entire module or file in one go, and it won’t lose track halfway through. Early feedback suggests it’s especially strong at explaining code, offering step-by-step justifications, and handling multi-step refactors that previously required multiple prompts.

In practice, your flow might look like this: you hit a bug, start a chat, and Auto routes you to Claude. The assistant remembers the earlier conversation, suggests a fix as a diff you can review in the editor, and explains why the change works. If you’re using the Claude Code extension instead, you can even let it apply changes across several files and run your test suite—all within the safety of your diff viewer.

Nevertheless, the model isn’t magic. It can still hallucinate APIs, misunderstand context, or generate subtly broken code. Human review, unit tests, and CI checks remain non-negotiable. And because Auto abstracts away the model choice, a small risk exists that the system picks a sub-optimal model for a niche question if another model is under heavy load. If you ever want absolute control, you can switch off Auto and choose a specific model from the Copilot dropdown.

Enterprise teams face additional considerations. Routing prompts to Anthropic’s servers—potentially via AWS Bedrock for some workloads—introduces cross-cloud data flows. Organizations with sensitive code must verify that their Copilot and Anthropic enterprise settings prevent data retention and training on proprietary code. Administrators can disable specific models from the admin portal if needed.

The Road to a Multi-Model Copilot

To understand why Microsoft made Claude the default, you have to look back a few months. In June 2025, an internal email from Julia Liuson, President of Microsoft’s Developer Division, told engineers that "based on internal benchmarks, Claude Sonnet 4 is our recommended model for GitHub Copilot." She cited superior performance in Microsoft’s own coding tests—guidance that remained even after OpenAI released GPT-5.

That internal verdict, later reported by multiple outlets, set the stage for the public launch of auto model selection in September. Behind the scenes, Microsoft had been evaluating Anthropic’s models alongside OpenAI’s and Google’s, part of a broader push to build a "multi-model moat." Mustafa Suleyman, co-founder of DeepMind and now a key figure in Microsoft’s AI efforts, has described the strategy as staying close to OpenAI while also investing in others and in Microsoft’s own models.

Anthropic, meanwhile, designed Claude 4 as two distinct models: Opus 4, the frontier coding model that scored 72.5% on the SWE-bench benchmark according to the company’s testing, and Sonnet 4, tuned for faster, cost-effective production use. The Copilot integration leans on Sonnet 4 because it balances capability and response time, which fits the interactive chat experience inside an IDE. Claude Code, by contrast, can tap into Opus 4’s full power for longer, agentic sessions where you’re willing to wait a bit longer for a more thorough answer.

The competitive context matters. GitHub Copilot originally wowed developers with GPT-3’s inline completions, but the coding AI market has exploded. Amazon CodeWhisperer, Google’s Gemini, and specialist tools like Cursor have all raised the bar. Microsoft’s answer is this orchestration layer: use the best model for each job, regardless of who builds it, and make the whole experience seamless enough that most developers won’t bother manually switching.

Getting Started with Claude 4 in VS Code

For most readers, the simplest way to start is already available if you have a paid Copilot subscription. Open VS Code, ensure the Copilot Chat extension is installed and signed in, and look for the model picker in the chat panel. Select "Auto"—that’s it. The next coding query you send will be routed according to Microsoft’s logic, which currently prioritizes Claude Sonnet 4. Hover over the response to confirm which model answered.

If you prefer the direct, agentic path, install Anthropic’s Claude Code CLI. You’ll need Node.js 18 or later, then run:

npm install -g @anthropic-ai/claude-code

Once you set your ANTHROPIC_API_KEY environment variable, you can open a terminal in any project and type claude. The IDE integration automatically installs and connects, giving you access to Opus 4, diff views, and multi-file edits. This route is ideal for experienced developers who want maximum control and don’t mind managing their own API costs.

A few precautions before you trust AI-generated changes:

  • Run unit tests after accepting any AI-written code.
  • Review diffs carefully—Claude’s suggestions can be wrong, especially with obscure libraries.
  • If you work with proprietary or regulated code, confirm your organization’s Copilot or Anthropic settings disable data retention.
  • For teams, enforce repository policies that require human approval on pull requests created by AI agents.

What’s Next for AI-Assisted Development

Microsoft has made it clear that auto model selection is just the start. The same architecture is being built into the full Visual Studio IDE and into Office 365 applications, where internal tests have already shown Claude outperforming GPT on certain Excel and PowerPoint tasks. As more models—including Google’s Gemini and perhaps open-source ones—join the lineup, the orchestrator will grow smarter about which model to use for which kind of request.

Anthropic is doubling down on agentic coding. Claude Code’s ability to run tests, create commits, and work for multi-hour sessions hints at a future where an AI pair programmer can take on larger, more autonomous chunks of software engineering. GitHub’s own “Copilot Agent” mode, powered in part by Claude, is in the pipeline, aiming to let the AI handle multi-step workflows directly in your repository.

For everyday developers, the message is clear: the AI behind your editor is no longer a single brand. It’s a committee, and for many coding tasks, that committee’s leader is now Claude. The shift promises better assistance, but it also demands a sharper eye for verification and governance. Those who learn to work with multiple AI models—knowing when to lean on Claude’s long-context reasoning versus, say, GPT’s rapid inline completions—will get the most out of this new multi-model world.