Model update (September 21): Grok 4.7 is rolling out in GitHub Copilot for Pro, Pro+, Max, Business and Enterprise. Organization administrators may need to enable it through model policy. This is Copilot access, not direct xAI API entitlement. The Grok 4.7 guide covers direct API pricing and the separate Cursor and Grok Build Fast option.

Sonnet update (September 28): Claude Sonnet 5.5 is generally available in Copilot for Pro, Pro+, Max, Business and Enterprise, with a gradual rollout. GitHub lists VS Code, Visual Studio, CLI, coding agent, Copilot App, github.com, Mobile, JetBrains, Xcode and Eclipse. Business and Enterprise administrators control access through model policy. See the Sonnet 5.5 guide for Anthropic API pricing and cloud availability.

Sol update (September 29): GPT-6.1 Sol is generally available with a gradual rollout for Copilot Pro+, Max, Business and Enterprise. GitHub lists VS Code, Visual Studio, Copilot CLI, coding agent, Copilot App, github.com, Mobile, JetBrains, Xcode and Eclipse. This is Copilot availability, not automatic direct API entitlement. See the GPT-6.1 Sol guide for the gpt-6.1-sol API version and migration differences from GPT-6 Sol.

GitHub Copilot is no longer just an IDE extension. At Microsoft Build 2026, Microsoft introduced a standalone GitHub Copilot App — a desktop application that brings AI coding assistance outside of VS Code and JetBrains. The app became generally available in June 2026 and is positioned as a workspace for agent-driven development across the full development lifecycle.

Architecture update (September 2026): VS Code 1.135 can discover supported recent Copilot sessions created in the Copilot CLI or standalone app and continue them through its persistent Agent Host. This is a supported-session workflow, not universal portability between every agent client. Our VS Code Agent Host and Agent Plugins 1.0 guide explains the boundary between session interoperability, harness behavior, skills, and MCP.

What’s new

Standalone desktop app

Copilot now exists as its own application, independent of any IDE. This means:

• Use Copilot without opening VS Code
• Plan features, discuss architecture, review code in a dedicated window
• Works alongside any editor (Vim, Neovim, Emacs, Sublime)
• Separate context and conversation from your IDE

Agentic development

Microsoft’s vision: multiple AI agents handling specific engineering tasks independently. The Copilot App supports:

• Planning agent — Break features into tasks, estimate complexity
• Implementation agent — Write code based on plans
• Debugging agent — Find and fix bugs
• Testing agent — Generate and run test suites
• Documentation agent — Write docs for your code

This is similar to Claude Code’s dynamic workflows but integrated into the Copilot ecosystem.

Model availability is not one fixed stack

GitHub’s current app announcement focuses on the workflow—isolated sessions, integrated terminal and browser, pull-request handoff, and Agent Merge—not on a permanent two-model stack. Available models and usage accounting can change by plan, so do not assume every Copilot App session runs one specific Microsoft model.

As of September 4, GPT-6 Astra is generally available specifically in GitHub Copilot for Pro+, Max, Business and Enterprise. GitHub lists it across VS Code, Visual Studio, Copilot CLI, coding agent, Copilot App, github.com, GitHub Mobile, JetBrains, Xcode and Eclipse. Business and Enterprise administrators control access through Copilot model policies. This Copilot availability does not establish general ChatGPT or OpenAI API availability.

GitHub removed four models from Copilot on October 2, 2026: Gemini 3.5 Flash and Gemini 3.6 Flash (replacement: Gemini 3.8 Flash), Kimi K2.7 Code (Kimi K3), and Claude Opus 4.7 (Claude Opus 5.5). They are no longer current Copilot choices. Managed organizations may need to enable replacements through model policy. This does not establish retirement from each vendor’s own API.

GitHub has also announced a separate retirement wave for October 19, 2026. This later change applies across Copilot experiences, including Copilot Chat, inline edits, ask and agent modes, and code completions.

Copilot model ending October 19	GitHub’s suggested replacement
Gemini 3.7 Flash	Gemini 3.8 Flash
GPT-5.5	GPT-5.6 Sol
GPT-5.4	GPT-5.6 Sol
GPT-5.4 mini	GPT-5.6 Luna
GPT-5 mini	GPT-5.6 Luna
Grok 4.5	Grok 4.6

For Business and Enterprise accounts using GitHub’s default model enablement, the suggested replacements are enabled automatically unless an administrator disabled the global default or explicitly blocked that model. Organizations with restrictive model policies should enable and test the replacement before October 19. GitHub removes the retired choices after the deadline, so administrators do not need a separate removal step. This is a GitHub Copilot lifecycle change only. It does not mean those models disappear from their vendors’ APIs or other products.

Auto routing: Efficiency, Balance and Intelligence

HydraFusion is orchestration, not Auto model selection

GitHub’s October 14 update keeps HydraFusion and HydraFusion Max in research preview across VS Code, Copilot App and CLI. GitHub lists Pro, Pro+, Business and Enterprise with preview features enabled, subject to organization policy. Auto routes a prompt to one model; HydraFusion coordinates models and stages inside the task.

Cascade can judge an initial result and request a repair only when needed. Critique inspects the actual workspace with tools and can run critics in parallel. The current pool includes Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, with models varying by workflow. Max uses the same architecture with deeper reasoning for demanding work. These are vendor-described techniques, not our independently reproduced performance results or a separately priced provider API model.

GitHub’s September 14 update introduces three auto-selection preferences, rolling out in VS Code, Copilot CLI and the Copilot App:

Preference	What it prioritizes	Practical starting point
Efficiency	Low cost	Straightforward edits or documentation
Balance	Cost, quality and response time together	Everyday development
Intelligence	Quality	Complex reasoning and implementation tasks

All three use the same available model pool. Auto evaluates each prompt, so Intelligence can still choose a small model for a simple request. Efficiency is not a fixed-price subscription or spending cap. Usage is charged for the actual selected model regardless of preference; paid subscribers retain GitHub’s documented 10% auto-usage discount. Availability still depends on rollout and model access. See Copilot usage-based billing before treating a routing preference as a predictable session budget.

How it compares to terminal coding tools

Copilot App’s unique position: it combines IDE-style features (autocomplete, inline suggestions) with agent-style features (autonomous task execution) in a single product. Terminal tools are more autonomous but lack the visual integration.

Desktop computer use: public preview

Since October 1, computer use is in public preview in Copilot CLI and the Copilot App on macOS and Windows. It can inspect accessible app content and visual context, click, type, use keyboard input, scroll and drag. This is GUI interaction, not an MCP server or a claim that every Copilot surface supports desktop control.

Enable it with /computer on; /computer show checks status and /computer off disables it. App users can also use Settings, Computer Use. Approval and OS permissions still apply, and organization policy can disable access. Do not infer Linux support from the App’s broader platform availability.

Copilot coding agent from Slack and Teams

Copilot’s cloud coding agent can also start from a team conversation. In Slack, mention @GitHub in a thread with a task; the agent works asynchronously and replies when a pull request is ready. The Teams integration follows the same collaboration pattern: the conversation supplies shared context, execution happens in a separate cloud environment, and the resulting change returns as a pull request for human review.

This is not the same as letting a chatbot edit the repository directly from Slack or Teams. The chat is the launch and coordination surface; GitHub remains the review and handoff surface. Developers can continue from the pull request in GitHub, an IDE, or a terminal. Availability depends on a paid Copilot plan and, for managed organizations, administrator policy. Usage and premium-request accounting can vary by plan, so teams should confirm billing before treating chat-originated tasks as an unlimited queue.

The practical workflow is: conversation context → asynchronous sandbox → branch and pull request → human approval. That separation is the important control boundary for collaborative agents.

Content exclusions in the Copilot App and CLI

As of September 2, 2026, the Copilot App and Copilot CLI generally respect content-exclusion policies configured by enterprise, organization, and repository administrators. This support is available to Copilot Business and Copilot Enterprise customers. Files matched by a policy are not used as direct context for Copilot responses in these clients.

The boundary matters. GitHub documents content exclusion as a context-control feature, not a complete secret-scanning or data-loss-prevention system. Semantic information may still reach Copilot indirectly through IDE features, and exclusions do not currently cover symbolic links or repositories on remote filesystems. GitHub also documents separate limitations for some editor agent modes. Teams should therefore combine exclusions with least-privilege credentials, secret scanning, repository permissions, and review controls rather than treating one policy as a universal protection layer. GitHub’s content-exclusion documentation has the current client matrix and limitations.

Enterprise-managed agent-operation permissions

Fresh policy and minimum client versions

The October 13 managed-settings update adds forceRemoteSettingsRefresh for Copilot CLI, VS Code and the App. When enabled, startup must confirm fresh server policy; a cached response is insufficient, and failure to retrieve current policy fails closed. MDM/file delivery can enforce this before first sign-in; server-delivered enforcement starts after the client has received the setting once. Without it, normal caching remains.

GitHub also begins minimum-client-version enforcement with the sandbox managed key. Older pinned clients must update to resume use. Policy freshness/version enforcement is not OS isolation: validate the effective filesystem, credential and network sandbox separately.

GitHub made centrally managed agent-operation permissions generally available on September 9 for Copilot Business and Enterprise. Administrators can classify shell commands, filesystem reads and edits, and network domains as denied, requiring human approval, or allowed. GitHub says a managed restriction cannot be weakened by a user’s workspace settings, auto-approval or previously saved approval.

The confirmed clients are the Copilot App, Copilot CLI and VS Code sessions using Agent Host. Policies can differ between enterprise teams. This is more specific than content exclusion: exclusions govern what files enter context, while operation permissions govern what an agent may do. Neither replaces repository permissions, secret isolation, sandboxing or pull-request review. See the AI Security foundation for the wider control model.

Local sandboxing is generally available

Copilot local sandboxing became GA on October 7 in CLI, the App and VS Code sessions using Agent Host, at no additional Copilot charge. Microsoft Execution Containers (MXC) applies filesystem, network and credential policy on Windows, macOS and Linux. This isolates tool execution, not the selected model’s inference. The September preview details below are historical; inspect the current effective policy rather than assuming every session enables sandboxing.

JetBrains compatibility

The October 10 plugin update ends support for IDE 2025.1: use 2025.2 or later. Enterprise-managed settings can choose a default agent model for new conversations while preserving the model picker. A separate MCP setting can disable automatic server startup.

Local sandboxing entered public preview in the Copilot App on September 23, 2026. It applies to locally running repository and working-tree sessions and is off by default. You can enable it for an active session with /sandbox on.

The policy can grant extra read/write paths, make paths read-only or deny them. Network controls cover outbound internet and local-network access, while separate switches govern Git HTTPS credentials and GitHub CLI authentication. Enterprise-managed policy can impose stricter limits than a project or user setting. GitHub says execution fails closed when the operating system cannot enforce the requested sandbox rather than silently continuing unsandboxed.

Local, cloud and remote-host sessions still need separate policy reviews. The former App-only preview description does not describe the full GA client coverage. Verify each environment’s filesystem, network and credential enforcement rather than treating GA as an automatic guarantee of isolation.

What this means for the competitive landscape

vs Cursor

Cursor is Copilot’s main competitor as an AI IDE. The Copilot App doesn’t replace VS Code/Cursor — it adds a companion window. But Microsoft making Copilot a standalone app signals intent to compete with Cursor at every layer.

vs Claude Code

Microsoft killed Claude Code licenses internally. The Copilot App with agentic capabilities is the replacement. For Microsoft-shop developers, this is the mandated tool going forward.

vs Aider/OpenCode

The Copilot App is proprietary and uses GitHub’s plan- and policy-dependent model catalogue, including third-party providers. Aider and OpenCode offer open-source alternatives with different supported-provider choices. Check the specific model and tool support rather than assuming any client works with every model.

Pricing and plan access

Gemini 4 Argon in Copilot

GitHub announced GA on October 13 for Pro+, Max, Business and Enterprise, with gradual rollout across VS Code, Visual Studio, CLI, coding agent, App, JetBrains, Xcode and Eclipse. Managed access follows model policies. Billing is provider-list usage pricing, not unlimited included model work. This Copilot entitlement is separate from direct Gemini API eligibility; see Gemini model selection.

GPT-6 Sol, GPT-6 Luna and Claude Opus 5.5

GitHub began a gradual September 22 rollout of three new model choices. GPT-6 Sol is available to Pro+, Max, Business and Enterprise; GPT-6 Luna also includes Pro; and Claude Opus 5.5 is available to Pro+, Max, Business and Enterprise. GitHub lists VS Code, Visual Studio, Copilot CLI, the coding agent, Copilot App, github.com, GitHub Mobile, JetBrains, Xcode and Eclipse as supported selection surfaces. Business and Enterprise administrators control access through model policies.

These are usage-billed model choices, not an increase in a plan’s included allowance. Rollout is gradual, so a documented model may not yet appear in every picker. For direct API economics and integration limits, see the GPT-6 Sol guide, GPT-6 Luna guide and Claude Opus 5.5 guide.

Enterprise OpenTelemetry export

The Copilot App now supports enterprise-managed OpenTelemetry export through the telemetry property in managed-settings.json. Administrators can send agent-session traces—including model and tool activity—to a compatible monitoring backend. Prompt and response content is excluded by default; enabling content capture is a separate governance decision.

This provides execution visibility, not automatic quality evaluation. Teams still need test sets and review gates to decide whether an agent produced the right result. Our OpenTelemetry monitoring guide explains the wider tracing model.

October 13 identity-control update: telemetry.capture.identity is off by default. Enabling it attributes sessions using user.name (GitHub login), process.user.name (local OS account) and host.name (machine), without requiring prompt capture. Administrators independently control prompts, responses, toolArguments, toolOutput and policyDetail; every capture class defaults off. Managed values override local settings. These controls apply to Copilot CLI, the app and VS Code Agent Host sessions. Raw content capture can expose code, commands, results and filesystem paths, so identity attribution and content collection require separate privacy decisions. GitHub’s October 12 announcement.

The generally available app is offered through eligible Copilot plans. Business and Enterprise access can depend on administrator and CLI policy settings. GitHub has not published a separate flat price for unlimited app or chat-originated agent work; included usage and premium-request accounting vary by plan.

When available

The standalone app has been generally available for Windows, macOS, and Linux since June 17, 2026. On Copilot Business or Enterprise, an organization or enterprise administrator must have Copilot CLI enabled in policy settings for app access.

What to use today

The Copilot App is generally available for eligible paid customers. Current options:

Want	Use today
IDE + AI (visual)	Cursor, current plan and usage charges apply
Terminal + autonomous agents	Claude Code or Aider
Free + open source	Aider + DeepSeek
Complex coding to evaluate	Claude Code with Opus 5.5; test task quality yourself
Microsoft/GitHub ecosystem	Copilot App or the Copilot extension in VS Code

See our best AI terminal coding tools for the full ranking.

FAQ

Is the Copilot App the same as Copilot CLI?

Different products. Copilot CLI is a terminal tool (already exists). The Copilot App is a standalone desktop GUI with agentic capabilities. Think of the App as the “desktop” version of the full Copilot experience.

Do I need VS Code to use the Copilot App?

No. That’s the point — it’s standalone. Use it with any editor (or no editor at all for planning/review tasks).

Will it replace the VS Code extension?

No. The extension continues for inline completions and editor integration. The App adds a separate window for complex interactions (planning, multi-agent, review).

Is it better than Claude Code?

There is no universal winner. Capability depends on the model, repository, task, governance requirements, and plan limits. Compare it with Claude Code on your own tests, review quality, and total usage cost.

Should I cancel Cursor and wait for this?

No. Both products are available now and serve overlapping but different workflows. Compare them on your repositories, governance requirements, review flow, and total usage cost before switching.

How does agentic development in the Copilot App differ from Claude Code dynamic workflows?

Claude Code dynamic workflows spawn hundreds of parallel subagents and generate JavaScript orchestration scripts. They are fully autonomous. The Copilot App’s agentic features appear more supervised — planning agents present plans for approval, implementation agents work under human review. This is more conservative but potentially safer for enterprise environments where full autonomy is risky.

Will this affect GitHub Actions and CI/CD?

Potentially. If Copilot agents can generate tests, run them, and fix code, the line between “developer tool” and “CI/CD pipeline” blurs. Microsoft has not announced direct GitHub Actions integration for the App yet, but the trajectory points toward agents that participate in the full PR → test → merge cycle autonomously.