GitHub Copilot adds Claude Haiku 5.5, local sandbox and stacked PRs GA simultaneously
GitHub released multiple Copilot updates on October 7: Claude Haiku 5.5 integration, local sandbox GA, stacked PRs GA, purpose-built model for leaked secret detection, CLI local model discovery.
On October 7, 2026, GitHub released multiple updates related to GitHub Copilot, spanning three directions: model integration, security protection, and collaboration tools.
On model integration, GitHub Copilot officially added Anthropic's Claude Haiku 5.5 model, in sync with Anthropic's October 7 launch. Claude Haiku 5.5 is Anthropic's fastest model to date, achieving 72.4% on the OSWorld 2.1 offline subset, a major improvement over Haiku 4.5's 15.7%; while maintaining intelligence level, pricing is on average 75% cheaper than Haiku 4.5, with 90% lower price for requests up to 100,000 tokens. After integration into GitHub Copilot, developers can choose claude-haiku-5-5 in coding conversation, completion, Chat, and Edits scenarios, balancing speed, cost, and capability.
On security, GitHub launched two key products. First, a purpose-built model for leaked secret detection, with higher accuracy in identifying sensitive strings like API keys, tokens, and certificates than general models, integrated into the secret scanning workflow. Second, GitHub Copilot local sandboxing became generally available (GA), providing a local isolated execution environment for Copilot-generated code and commands. Local sandboxing was originally launched as a research preview, intended to prevent Copilot tool calls from accessing sensitive resources or performing unintended actions; after GA, it is enabled by default for all users and supports custom sandbox policies.
On collaboration tools, GitHub's Stacked Pull Requests (Stacked PRs) feature became generally available (GA). Stacked PRs let teams efficiently manage interdependent PR chains, avoiding the traditional "PR chain mess" problem: when one PR depends on another unmerged PR, developers can organize them in a stack relationship, with independent review, testing, and merging. This feature was released to GA on October 6, and on October 7 reached a broader user base.
In addition, GitHub Copilot CLI added the "Discover local models" feature, letting users find and configure locally available AI models in the command line, including local runtimes like Ollama. This feature lets developers scan locally installed models and use them as Copilot backend alternatives without leaving the CLI.
Overall, GitHub's four key updates on October 7 continue to strengthen Copilot's leadership in the coding agent market across four dimensions: model diversity (multi-vendor support), execution security (sandbox), code security (secret detection), and collaboration workflow (stacked PR). Combined with the previously released Copilot computer use public preview, these updates together represent a key step in GitHub's evolution from "code completion tool" to "full-stack AI coding partner".