The latest AI news and updates of GitHub — Developer platform powering GitHub Copilot and AI-assisted software development workflows. Covering GitHub's latest product updates, launches, and company news from the past 90 days.
GitHub21h agoGitHub made OpenAI's GPT-6.1 Terra generally available in Copilot for everyday agentic and terminal work. Early testing shows the model solves more terminal and C# coding tasks than GPT-5.6 Terra, operating in focused, consolidated steps. It is available to Pro, Business, and Enterprise users at a lower cost than GPT-6.1 Sol, with gradual rollout across all Copilot surfaces.
GitHubOct 18GitHub Security Lab used its open-source Taskflow Agent to identify 24 Android vulnerabilities, including a critical location leak in a navigation app with 10 million downloads. The lab published the Android-specific audit taskflows for running these security scans on local repositories. The process requires a GitHub Copilot license and involves premium model requests.
GitHubOct 13GitHub made Google's Gemini 4 Argon generally available in GitHub Copilot for complex coding, agentic, and reasoning tasks. Early testing shows it solves more tasks than previous Gemini models on every coding benchmark, with largest gains in Copilot CLI. It thoroughly checks work and tests edge cases before finishing. Available to Pro+, Max, Business, and Enterprise users.
GitHubOct 11GitHub Security Lab released the Fuzzing Taskflow, an autonomous pipeline for C and C++ projects. The agent identifies entrypoints, analyzes builds, writes harnesses, runs AFL++, and triages crashes into unique vulnerability reports. It automates the entire continuous fuzzing loop, including coverage-feedback and crash deduplication, to eliminate manual oversight in security testing.
GitHubOct 8GitHub launched a context-aware classifier that evaluates candidate secrets in under 2 milliseconds. Built on a ModernBERT architecture, the model could more than double the number of secrets blocked by push protection before they reach repository history. The classifier is currently in private preview and is now integrated into the Copilot CLI and App security review commands.
GitHubOct 8GitHub made Anthropic's Claude Haiku 5.5 generally available in GitHub Copilot. Designed for high-volume tasks like subagents and terminal work, the model matches Claude Sonnet 5 performance on many coding tasks while using fewer tokens and steps. It is available to Pro, Pro+, Max, Business, and Enterprise users across all Copilot surfaces, with usage-based billing.
GitHubOct 8GitHub launched local sandboxing for GitHub Copilot, providing an isolated execution boundary for agentic workflows. The feature restricts access to files, networks, and credentials across the Copilot CLI, Copilot app, and VS Code Agent Host. Centralized enterprise policies enforce these boundaries, ensuring secure tool execution at no additional cost.
GitHubOct 8GitHub will soon update Copilot to automatically route tasks between local and cloud models, reducing AI credit usage. This Project HydraFusion update includes sandboxed tool execution via Microsoft Execution Containers. The feature supports local inference on NVIDIA RTX Spark Windows PCs using a quantized version of the MAI Code 1.1 Flash model, available by the end of the month.
GitHubOct 7GitHub is rebuilding its Git infrastructure to support agent-scale software development. With Git activity doubling to 473.3 billion events per month, the platform is decoupling storage from compute and moving maintenance off the serving path. Internal benchmarks show the new architecture delivers up to 35 times higher write throughput while maintaining existing governance and control.
GitHubOct 6GitHub launched ReviewBench, an open benchmark for evaluating AI code review agents. Built from 103.9 million GitHub pull requests, the dataset includes 219 PRs across 19 languages. It provides a multi-source golden set and structured findings to measure precision and recall, allowing teams to bring their own agents, evaluate performance, and publish results on a common leaderboard.
GitHubSep 29GitHub made OpenAI's GPT-6.1 Sol generally available in GitHub Copilot for agentic coding and terminal workflows. Early testing shows the model completes tasks using fewer tokens and steps than previous GPT-6 and GPT-5.6 models. It is available to Pro+, Max, Business, and Enterprise users across the Copilot app, CLI, and supported IDE extensions.
GitHubSep 28GitHub added Anthropic's Claude Sonnet 5.5 to GitHub Copilot for everyday coding tasks like building features and fixing bugs. Early testing shows the model matches Sonnet 5 performance while using significantly fewer steps, tokens, and tool calls, resulting in faster task completion. The model is available across the Copilot app, CLI, and supported IDE extensions.
GitHubSep 26GitHub argues that chat is often the wrong interface for AI-powered development, proposing canvases as a more effective alternative. Canvases function as full-stack applications running inside the GitHub Copilot app, allowing bidirectional communication with the agent and local code execution. This surface supports custom, task-specific tools, replacing the generic chat box for complex workflows.
GitHubSep 24GitHub rebuilt the pull request view in the GitHub Copilot app to maintain performance on massive diffs. The surface handles a pull request with 2,200 files, over a million changed lines, and 400 inline comments by separating code and comment geometries. Idle-gated measurement and scroll anchoring keep the interface responsive during complex reviews.
GitHubSep 22GitHub added OpenAI’s GPT-6 Sol and GPT-6 Luna models to GitHub Copilot, joining the previously released GPT-6 Astra. GPT-6 Sol provides a balanced model for interactive and agentic coding, while GPT-6 Luna serves as the family’s lightweight, lowest-cost option. Both models are generally available with usage-based billing, with access managed through Copilot settings.
GitHubSep 22GitHub added Anthropic's Claude Opus 5.5 to GitHub Copilot for Pro+, Max, Business, and Enterprise users. Early testing shows the model resolves tasks comparably to Claude Opus 5 while using significantly fewer steps and tokens, with faster error recovery on multistep tasks. The model is available across the Copilot app, CLI, and supported IDE extensions.
GitHubSep 21GitHub ported the GitHub Copilot agent runtime to Rust, replacing over 800,000 lines of production TypeScript. One engineer and a team of agents completed the migration in 14.5 weeks using an incremental, in-place strategy. The new native runtime supports in-process hosting for SDKs, significantly reducing memory overhead and startup latency while maintaining existing agentic capabilities.
GitHubSep 15GitHub added the My work pane to the GitHub Copilot app, providing a centralized hub to track pull requests and issues. The pane supports custom views, repository filtering, and the ability to initiate new agent sessions directly from specific tasks. It organizes development workflows by grouping active, pending, and completed items in one location.
GitHubSep 9GitHub Copilot code review now submits formal pull request approvals, addressing the 84% of PR lifecycle time spent waiting for sign-off. Administrators can enable this feature at the enterprise, organization, or repository level, with approvals counting toward required-approvals rules. Copilot automatically dismisses approvals if new commits are pushed, and the feature is currently in public preview.
GitHubSep 8GitHub updated its Copilot agent harness to reduce token waste without sacrificing task quality. By selectively compressing output, removing redundant line-number formatting, and batching background notifications, the system avoids unnecessary model turns. A meta-prompting loop also reduced prompt tokens by 1.8% per session. These changes prioritize task completion over minimizing individual tool-call tokens.
GitHubSep 6GitHub introduced canvases in the Copilot app to provide a durable, shared surface for agentic workflows. Canvases replace long chat scrolls by tracking what ran, changed, and requires human review. The feature includes example templates like Java Modernization Studio and Site Studio, available in awesome-copilot, to maintain state and governance during complex, multi-step AI tasks.
GitHubSep 4GitHub introduced Project HydraFusion, a research preview that dynamically orchestrates multiple AI models to execute coding tasks. In offline evaluations on Terminal-Bench 2.1, the system improved verified task quality by 4.9 percentage points while reducing estimated costs by 67% compared to Claude Opus 5. The preview is available through the /experimental flag in the GitHub Copilot CLI.
GitHubSep 3GitHub added Google's Gemini 3.8 Flash to Copilot, now available in the app, CLI, and VS Code. The model shows strong results on complex terminal-based coding tasks. Access extends to Pro, Pro+, Max, Business, and Enterprise plans with introductory provider pricing through December 31, 2026.
GitHubSep 2GitHub added Anthropic's Claude Fable 5.1 to GitHub Copilot for Pro+, Max, Business, and Enterprise plans. The model performs on long-running coding tasks, deep codebase research, and complex agentic workflows. It requires data retention by default for safety classifiers, though eligible enterprises can access zero-data-retention endpoints under a time-bound exception.
GitHubAug 22GitHub is offering a 30% discount on auto model selection for Copilot Max users. This feature automatically routes tasks to Claude, GPT, or Microsoft AI models without requiring manual code or opt-in. The promotional pricing remains active until 00:00 UTC on September 30, 2026.
GitHubAug 16GitHub released Lite and Balanced effort levels for Copilot code review, replacing the previous Low and Medium settings. Balanced provides deeper analysis for complex pull requests, while Lite targets straightforward changes. Organization-level defaults govern review depth across repositories, with per-request selection available for Pro, Pro+, Max, Business, and Enterprise plans.
GitHubAug 13GitHub is rolling out Google's Gemini 3.7 Flash model to GitHub Copilot. The model improves web and app development, agentic coding workflows, and codebase research. It is available to Copilot Pro, Pro+, Max, Business, and Enterprise users across IDEs including VS Code, JetBrains, and Xcode. Business and Enterprise administrators must enable the model policy to grant access.
GitHubAug 11GitHub is rolling out Microsoft's MAI-Code-1.1-Flash model to Copilot, featuring native vision support for image understanding and improved coding performance. The update delivers a 73% lower list price compared to the previous version. The model is available across Copilot interfaces, with manual selection options for Pro and Enterprise plans and auto-selection for Free and Student users.
GitHubAug 6GitHub added Moonshot AI’s Kimi K3, an open-weight model, to GitHub Copilot. The model features frontier-level agentic coding capabilities and is hosted on Fireworks AI infrastructure. It is rolling out to all paid plans, though Business and Enterprise administrators must manually enable the model policy. Pricing follows provider list rates, starting at $3 per million input tokens.
GitHubAug 5GitHub added voice prompting to the GitHub Copilot desktop app, replacing manual typing with spoken input. A new microphone icon in the app interface captures and transcribes speech directly into the session. This feature is available on macOS, Windows, and Linux for all Copilot plans.
GitHubAug 3GitHub’s Aspire team implemented GitHub Agentic Workflows to automate cross-repo documentation. The pipeline generates draft pull requests from merged features, requiring subject-matter expert review. The team merged 82 documentation pull requests with a 44.8-hour median time-to-merge, keeping documentation synchronized with code changes through automated intent-based processing.
Frequently asked questions