Three companies I advise asked me the same question. One CTO put it bluntly: “We already pay for Copilot. Someone told me Claude is in there now. So why would I also pay for Claude separately?”

Fair question. The answer isn’t simple, and the vendors aren’t making it any easier to understand.

Claude models are available inside GitHub Copilot as of February 2026. That’s real. Since I wrote this, Claude also landed in mainline Microsoft 365 Copilot chat through Microsoft’s Frontier program, so the OpenAI-only days are over there too. But what you get through Copilot is not the same experience as what you get from Anthropic directly. Same engine, different car. The controls, the dashboard, the range, the storage space are all different.

Same model, different box

Think of it like Spotify running inside a car’s infotainment system versus the Spotify app on your phone. Same music library. But in the car, you can’t browse playlists the same way, the interface is clunkier, and some features just aren’t available. You’re still listening to the same songs, sort of. The experience is fundamentally different.

That’s what’s happening with Claude inside Copilot.

GitHub’s model picker now includes Claude Fable 5, Opus 4.8, Opus 4.7, Sonnet 4.6, Sonnet 4.5, and Haiku 4.5. You select which one you want in Copilot Chat, and the requests route to that model. Copilot stopped being an OpenAI-only product a while back. Today it spans OpenAI, Anthropic, Google, and Microsoft’s own models, so the question is which Claude you pick, not whether Claude is there.

This happened because Microsoft and Anthropic struck a massive Azure infrastructure deal. Claude is now available through Microsoft Foundry, which powers Copilot’s multi-model support. Amazon Bedrock offers a parallel delivery channel for enterprises on AWS. The pattern that keeps showing up: every cloud provider wants Claude running inside their ecosystem, on their terms.

Here’s the bit that matters for procurement decisions. When your VP of Engineering says “we have Claude in Copilot,” they’re technically correct. When your senior architect says “that’s not really Claude,” they’re also correct.

Both are right. Both are talking past each other.

What Copilot gives you for free

If your company already pays for GitHub Copilot Business or Enterprise, your developers can start using Claude models today without any additional subscription. Zero extra cost. That’s useful and I’d be a proper fool to dismiss it.

The model picker in VS Code takes about ten seconds to configure. Click the model name in the chat input, select “Manage Models,” expand the Copilot section, pick your Claude variant. Done. On the web at copilot.github.com, it’s even simpler: dropdown arrow in the prompt box, select Claude.

IT admins control this at three levels. Enterprise-wide policy. Organization-level settings. Individual user activation. That layered governance is one of Copilot’s real strengths for mid-size companies. No new vendor relationship means no new security review, no new procurement cycle, no new data processing agreement. For a 200-person company where the security team is already stretched thin, this is a no-brainer.

Claude through Copilot works in three surfaces: Copilot Chat on the web, the VS Code extension, and as a coding agent. For teams that live inside VS Code all day and just need a smarter autocomplete and chat assistant, this covers a lot of ground.

Worth it to talk about your specific shape of this? Blue Sheen is set up for that.

What you lose inside Copilot

This is where it gets interesting. And a bit annoying, if I’m being straight.

The context window shrinks. Claude’s standalone context supports up to 1M tokens on current models like Sonnet 4.6 and Opus 4.6 and up. Through Copilot, community reports and documentation indicate a ceiling around 128,000 to 150,000 tokens. For a small project, you won’t notice. For a system with 40 interconnected files and complex business logic, the difference between a tool that holds your entire codebase in its head versus one working with fragments is night and day.

Update (June 2026): GitHub has since raised Copilot’s context ceiling. Pick a current Claude model inside Copilot (Sonnet 4.6, or Opus 4.6 and up) and you now get up to 1M tokens, so this particular gap has mostly closed. The extended-thinking and MCP gaps further down still stand.

Extended thinking disappears. Claude’s ability to reason through multi-step problems before answering, to essentially think out loud internally, to take time before responding to complex architecture questions, that’s not available through the Copilot interface. No adjustable reasoning budget. No thinking tokens. You get the model’s first-pass response, which is still good. But it’s not the same as what you get when the model takes 30 seconds to think through something difficult.

No MCP support. The Model Context Protocol lets Claude connect directly to enterprise tools: Jira, Confluence, Slack, databases, internal APIs. Inside Copilot, Claude sees what Copilot feeds it, and nothing more. That means no pulling context from your project management tool mid-conversation. No querying your staging database to check a migration. The model is brilliant but blind to the rest of your tooling.

No Projects, no persistent memory, no conversation history that carries over between sessions. Every Copilot interaction starts fresh. Claude’s standalone Projects feature lets you build persistent workspaces with instructions, knowledge files, and memory that accumulates over time. Through Copilot? Each chat is a blank slate.

Rate limits work differently too. Copilot allocates a shared pool of “premium requests” across advanced models. Heavy Claude users will hit ceilings that don’t exist on a dedicated Claude subscription. I watched one team burn through their monthly premium allocation in the first week, then get downgraded to a lighter model for the remaining three weeks. That’s a painful way to discover the limitation.

The pricing reality for mid-size companies

I won’t quote specific prices because they change constantly, but the structure matters.

Copilot has a tiered subscription: a free tier with limited completions and requests, individual pro plans for solo developers, a plus tier for power users who need more premium model access, a business tier adding organizational governance, and an enterprise tier with compliance features and customization. Claude access is bundled into these tiers at no additional per-model charge.

Claude has its own tiers: a free conversational tier, a Pro subscription for regular users, a Max tier for heavy usage, and direct API access billed per token.

The question I get asked most often: “Aren’t we paying twice?”

Not exactly. Copilot Business gives you “some Claude” included. A dedicated Claude subscription gives you “all of Claude” separately. They’re covering different surfaces of the same problem. One lives in your IDE. The other lives everywhere else. Running both isn’t doubling your spend; it’s covering two distinct use cases that overlap in a thin middle band.

For a team of 50 engineers, the math changes based on how many actually need deep Claude capabilities versus how many are perfectly served by Copilot’s version. In conversations I’ve had with similar-sized companies, the split usually lands around 80/20. Eighty percent of the team is fine with Copilot’s Claude. Twenty percent need the real thing for architecture work, security reviews, and complex debugging sessions.

Tracking actual usage patterns before committing to additional seats saves you from buying licenses that sit unused.

When you need both

The pattern that keeps emerging is straightforward. Developers use Copilot for inline suggestions, quick completions, code review assistance, and fast chat-based questions inside their editor. Stays in the flow. Keeps the cursor moving. That’s Copilot’s sweet spot and it does it brilliantly.

Those same developers switch to Claude directly for architecture planning, complex multi-file refactoring, writing documentation from scratch, debugging production issues that span multiple services, or anything requiring extended context. These tasks benefit from the full 1M-token window, extended thinking, and MCP integrations that pull context from the rest of the toolchain.

This isn’t a theory. It’s a multi-model routing pattern that shows up everywhere once you look for it. Different models and interfaces for different cognitive loads. Quick generation goes to the fast, embedded tool. Deep reasoning goes to the full-featured one.

The governance approach that works for most mid-size companies: enable Claude models inside Copilot for every developer immediately. It costs nothing additional if you’re already on Copilot Business. Then provision dedicated Claude Pro or Max seats only for the developers and tech leads who demonstrate real need for the full capabilities. That’s usually architects, security engineers, team leads doing design reviews, and anyone working on legacy modernization or complex integrations.

If you don’t provide Claude access through approved channels, developers get it themselves through personal accounts. That’s not a hypothetical. It’s happening at every company I talk to. Better to offer it officially through Copilot first, then add standalone seats where the demand proves itself.

Two questions to answer, and then the decision basically makes itself. Does your team already use Copilot? If yes, enable Claude models inside it today. Takes 10 minutes. Do any of your teams regularly need extended thinking, MCP connections, or context windows above 150,000 tokens? If yes, add dedicated Claude seats for those specific people. Nobody else needs them.

The confusion between Claude and Copilot is understandable. The vendors benefit from it. Turns out the answer isn’t choosing one. It’s using each where it belongs.