What Happened
GitHub expanded Copilot’s model choices for coding and review. GPT-6.2 Astra and Sol are generally available to eligible paid users; Microsoft’s MAI Code 2 and lighter Code 2 Flash are also available, with Flash extending to Free and Student users. Enterprise administrators can control model access through policies. GitHub also added models to Copilot Fusion and reported a 71.2% resolution rate for Fusion Max on Terminal-Bench 4.1. That benchmark is a useful comparison, not a prediction of performance on a company’s repositories. [1][4][7]
Copilot code review gained public previews of risk-based Adaptive effort, cross-repository context for public repositories, and secret detection. Review for Azure Repos is generally available. Repository insights and usage-metrics reports now show whether comments were addressed and where review coverage is missing; Copilot Evals, in public preview, tests both agent results and the steps taken to reach them. [6][11][12]
The Copilot app added a public-preview Azure DevOps integration and opt-in browser use, while GitHub Enterprise Server support is generally available in app version 1.2.0 or later. JetBrains gained preview cloud sessions, cross-client access, browser use, and voice dictation, alongside remote-development and reliability changes. GitHub’s Agent Host Protocol supports reconnectable sessions across clients, including VS Code and cloud sandboxes. Enterprise Managed Users can stream agent activity to Microsoft Agent 365, and Copilot CLI added preview Microsoft Entra sign-in on Windows and macOS. [3][5][8][9][10]
Why It Matters to Businesses
The purchase decision is shifting from “Which model writes code fastest?” to “Can agents find the right context, produce reviewable changes, and operate within enterprise controls?” This is especially important in monorepos, where a local checkout may omit code an agent needs; searchable, precisely navigable repository context matters more than simply increasing a prompt’s size. The notes establish new GitHub capabilities and integrations, but not comparable new releases from Cursor, Windsurf, Replit, Sourcegraph, or Continue.dev. [13]
More findings do not automatically mean better security. In a six-project pilot, GitHub and Microsoft reduced thousands of potential vulnerability findings to 99 privately reported, reviewed findings with suggested fixes. Maintainers had fixed 55 by November 15, but researcher review took roughly two days per project. The operational lesson is to budget for verification and remediation, not just scanning. [2]
Kimbodo Engineering Perspective
We would treat model selection, repository retrieval, agent execution, and code review as separate controls. A cheaper model may suit routine edits, while a stronger model may justify its cost for multi-file changes or high-risk review. Adaptive effort can help allocate review capacity, but teams should validate its decisions against their own pull requests and keep human approval for consequential changes. Preview features warrant narrower rollout than generally available ones. [1][6][7][12]
How We Would Implement It
- Start with a representative set of bugs, refactors, and review cases. Record correctness, human rework, latency, token or usage cost, and missed high-severity issues; rerun Copilot Evals after changing models, instructions, skills, or MCP tools. [1][7][12]
- Give agents indexed code search and explicit repository boundaries. For monorepos, test whether retrieval finds dependencies outside sparse checkouts before enabling broad autonomous edits. Restrict cross-repository context to approved material. [6][13]
- Require branch-based changes, CI checks, secret scanning, and human pull-request approval. Track addressed review comments and unreviewed merges in Repository insights rather than counting generated comments as success. [6][11]
- Roll out authentication, model policies, and audit streaming before cross-device sessions or browser use. Enable browser use only for defined test workflows and approved accounts; pilot Entra CLI sign-in only where its current platform and license requirements fit. [3][5][9][10]
Risks, Costs and Security
Usage-based model pricing, Azure Repos review credits, and higher-effort reviews can make costs variable; measure cost per accepted change, not cost per prompt. Browser automation operates in signed-in sessions, and persistent agent sessions increase the importance of scoped credentials, session revocation, and audit review. Secret detection is a preview safeguard, not a replacement for established secret scanning and incident response. For externally reported vulnerabilities, verify exploitability and coordinate private, actionable fixes with maintainers. [1][2][5][6][9]
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our AI Application Development practice, or Estimate My AI Application.
Sources
- [1] OpenAI’s GPT-6.2 Astra and GPT-6.2 Sol now available
- [2] Maintainers don’t need more AI bug reports. They need fixes.
- [3] Building the Agent Host Protocol for sessions that continue across devices
- [4] GitHub Copilot Fusion updates: more availability, benchmarks, and models
- [5] GitHub Copilot app, Microsoft Ignite releases
- [6] Copilot code review: Adaptive effort, secret detection, and more
- [7] MAI Code 2 and MAI Code 2 Flash in GitHub Copilot
- [8] GitHub Copilot in JetBrains, Ignite releases
- [9] Monitor GitHub Copilot activity through Microsoft Agent 365
- [10] GitHub Copilot CLI adds Microsoft Entra sign-in as a preview
- [11] Track Copilot code review outcomes in Repository insights
- [12] GitHub Copilot Evals is now in public preview
- [13] Why large monorepos need code search, especially with AI coding agents