GitHub Copilot
Last reviewed
May 13, 2026
Sources
50 citations
Review status
Source-backed
Revision
v8 ยท 8,747 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 13, 2026
Sources
50 citations
Review status
Source-backed
Revision
v8 ยท 8,747 words
Add missing citations, update stale details, or suggest a clearer explanation.
GitHub Copilot is an artificial intelligence (AI) coding assistant developed by GitHub and OpenAI. It provides autocomplete-style code suggestions, chat-based assistance, and autonomous agent capabilities directly within a developer's integrated development environment (IDE). Since its launch as a technical preview on June 29, 2021 and general availability on June 21, 2022, Copilot has grown into one of the most widely adopted AI-powered developer tools in the world, surpassing 20 million total users by mid-2025 and generating over $2 billion in annual recurring revenue for Microsoft.
The tool was the first major commercial product to emerge from Microsoft's $1 billion investment in OpenAI. It analyzes the context of code being written, including comments, function names, and surrounding code, and generates suggestions ranging from single-line completions to entire functions and classes. Originally powered by OpenAI Codex (a descendant of GPT-3), GitHub Copilot has evolved into a multi-model platform that supports models from OpenAI, Anthropic, Google, and xAI. The product has also been the subject of a high-profile class action lawsuit (Doe v. GitHub) over the use of open-source code in its training data, and has driven a substantial body of academic and industry research into the productivity effects of AI coding assistants.
GitHub Copilot operates as an extension or plugin inside supported IDEs. As a developer types, the tool sends context (the code in the current file, related open files, comments, file names, and import statements) to a hosted large language model. The model returns one or more candidate completions, which the developer can accept with the Tab key, cycle through, or dismiss. The product has expanded well beyond inline completion to include a conversational chatbot interface (Copilot Chat), multi-file refactoring (Copilot Edits), an autonomous agent that can edit code, run commands, and iterate on test results (agent mode), an asynchronous cloud agent that can be assigned GitHub issues (the coding agent), and a code review feature that comments on pull requests.
Four features account for most of Copilot's value to professional developers. The first is inline code completion, the original feature from 2021 that still drives the bulk of suggestions accepted in the wild. The second is Copilot Chat, an in-IDE conversational interface that became generally available in December 2023. The third is multi-file editing, introduced in late 2024 as Copilot Edits, which lets a developer describe a change in natural language and have Copilot apply it across a working set of files. The fourth is the agent layer, including the in-IDE agent mode announced in February 2025 and the asynchronous coding agent announced at Microsoft Build in May 2025 and made generally available later that year, which can take a GitHub issue, write the code, run tests in a cloud environment, and open a pull request without further human intervention.
GitHub Copilot's history runs from a small research collaboration between GitHub and OpenAI through a paid SaaS product and finally into a multi-model agent platform. The table below summarises the major launches.
| Date | Event |
|---|---|
| June 29, 2021 | Technical preview announced by GitHub CEO Nat Friedman; available as a VS Code extension |
| October 27, 2021 | Neovim plugin released as a public repository |
| October 29, 2021 | JetBrains marketplace plugin published |
| March 29, 2022 | Support for Visual Studio 2022 announced |
| June 21, 2022 | General availability for individuals at $10 per month or $100 per year |
| September 2022 | Free access announced for verified teachers on GitHub Global Campus |
| November 3, 2022 | Doe v. GitHub class action lawsuit filed in the Northern District of California |
| December 7, 2022 | Free access expanded for verified students; Copilot for Business beta begins |
| February 14, 2023 | Copilot for Business reaches general availability at $19 per user per month |
| March 22, 2023 | Copilot X vision announced, including Copilot Chat, CLI, Docs, and Pull Requests features powered by GPT-4 |
| July 20, 2023 | Copilot Chat enters public preview for businesses |
| November 8, 2023 | Copilot Enterprise teased at GitHub Universe with knowledge bases trained on customer code |
| December 29, 2023 | Copilot Chat reaches general availability |
| February 27, 2024 | Copilot Enterprise reaches general availability at $39 per user per month |
| April 29, 2024 | Copilot Workspace technical preview announced |
| May 21, 2024 | Copilot Extensions enter limited public beta |
| October 29, 2024 | Multi-model support announced at GitHub Universe 2024 with Claude 3.5 Sonnet, Gemini 1.5 Pro, and OpenAI o1-preview |
| October 29, 2024 | GitHub Spark previewed as a natural language micro-app builder |
| November 12, 2024 | Copilot Edits preview launches in VS Code |
| December 18, 2024 | GitHub Copilot Free tier launches with 2,000 completions and 50 chat messages per month |
| February 6, 2025 | Agent mode and Next Edit Suggestions announced |
| April 2025 | Agent mode reaches general availability |
| May 19, 2025 | Coding agent announced at Microsoft Build |
| May 30, 2025 | Copilot Workspace technical preview sunset |
| July 2025 | Total users pass 20 million |
| September 2025 | Coding agent reaches general availability; Copilot Extensions deprecated in favour of MCP |
| Late 2025 / early 2026 | Paid subscribers reach 4.7 million |
The foundations of GitHub Copilot trace back to the development of the GPT family of large language models at OpenAI. In 2018, OpenAI released the first version of GPT with 110 million parameters, demonstrating that transformer-based models could generate coherent text. The following year, GPT-2 was released, trained on 40 GB of internet text with 1.5 billion parameters. Its text generation capabilities raised questions about potential applications in code generation, and several research groups, including those at GitHub, began experimenting with the model on programming tasks. In 2020, GPT-3 was announced with 175 billion parameters, trained on a large corpus that included source code from public repositories. Researchers at OpenAI began exploring a code-specialised variant, which eventually became Codex. Microsoft, which had acquired GitHub in 2018 for $7.5 billion and made a $1 billion investment in OpenAI in 2019, was uniquely positioned to combine the platforms and develop a commercial product.
On June 29, 2021, GitHub announced GitHub Copilot as a technical preview, available exclusively as a Visual Studio Code extension. The announcement was made by then GitHub CEO Nat Friedman in a blog post titled "Introducing GitHub Copilot: your AI pair programmer." The tool was positioned as an AI pair programmer that could assist developers with real-time code suggestions. At launch, GitHub reported that roughly 30% of its suggestions were accepted by testers as correct and bug free.
The technical preview was powered by OpenAI Codex, a model derived from GPT-3 but fine tuned on billions of lines of public source code from GitHub repositories. The Codex paper, Chen et al. "Evaluating Large Language Models Trained on Code" (arXiv:2107.03374), described training on a filtered 159 GB dataset of Python from 54 million public repositories, with strong performance also on JavaScript, TypeScript, Ruby, Go, and a handful of other languages. The technical preview was free to join via a waitlist, and hundreds of thousands of developers signed up in the weeks after the announcement.
Additional IDE support followed within months. On October 27, 2021, GitHub released the Copilot plugin for Neovim as a public repository. On October 29, 2021, the Copilot plugin was published on the JetBrains marketplace, bringing the tool to IntelliJ, PyCharm, GoLand, and other JetBrains IDEs. On March 29, 2022, GitHub announced Copilot support for Visual Studio 2022.
On June 21, 2022, GitHub Copilot was made generally available to all individual developers through a paid subscription at $10 per month or $100 per year. Verified students and maintainers of popular open-source projects received free access. The launch was a significant commercial milestone: it was the first time a major AI vendor had charged a monthly subscription for a generative AI product. Within months, GitHub reported strong revenue growth driven by Copilot, and the product was credited with accelerating Microsoft's broader push into AI.
In September 2022, GitHub extended free access to verified teachers on GitHub Global Campus, with the explicit goal of bringing Copilot into computer science classrooms. In December 2022, GitHub announced Copilot for Business as a paid beta at $19 per user per month, adding organisational management features, security policy controls, and IP indemnity. Copilot for Business reached general availability on February 14, 2023, with GitHub reporting more than 400 organisations already enrolled at the time of launch.
In March 2023, GitHub announced Copilot X, a vision for the next generation of the tool. Copilot X introduced several new capabilities powered by GPT-4:
In July 2023, Copilot Chat entered public preview for organisations on the Business plan. The chat experience was rolled out gradually, with individual subscribers gaining access later in the year. In November 2023, Copilot Chat was updated to use GPT-4 as its default model, replacing earlier GPT-3.5 based versions, and on December 29, 2023, Chat reached general availability.
At GitHub Universe 2023 (November 8, 2023), GitHub also previewed Copilot Enterprise, which added features like knowledge bases trained on an organisation's own codebase, Copilot Chat on GitHub.com, and integration with internal documentation. Enterprise reached general availability on February 27, 2024, at $39 per user per month, with the ability to attach knowledge bases (formerly known as docsets) to chat conversations, generate pull request summaries, and chat about changes in pull requests.
On April 29, 2024, GitHub launched a technical preview of Copilot Workspace, an experimental browser-based development environment that could turn a GitHub issue into a specification, a plan, and a set of code changes through a sequence of sub-agents. On May 21, 2024, Copilot Extensions entered limited public beta with partners including DataStax, Docker, LambdaTest, LaunchDarkly, McKinsey & Company, Microsoft Azure and Teams, MongoDB, Octopus Deploy, Pangea, Pinecone, Product Science, ReadMe, Sentry, and Stripe.
At GitHub Universe 2024 on October 29, 2024, GitHub announced two major shifts. The first was multi-model support: developers could now choose between Anthropic's Claude 3.5 Sonnet, Google's Gemini 1.5 Pro, and OpenAI's GPT-4o, o1-preview, and o1-mini for Copilot Chat. The second was GitHub Spark, a natural language app builder for small web applications. GitHub CEO Thomas Dohmke described the change in a keynote, arguing that there is no one model that suits every scenario and that the next phase of AI code generation will be defined by multi-model choice. Claude 3.5 Sonnet was rolled out via Amazon Bedrock, using cross-region inference to improve reliability.
On November 12, 2024, GitHub previewed Copilot Edits, a multi-file editing experience that combined the conversational flow of Chat with inline edits across a working set of files. Edits introduced a UI where the developer could drag files into a working set, describe a change in natural language, and review the resulting diffs across multiple files at once.
On December 18, 2024, GitHub launched a free tier for Copilot, making the tool accessible to any developer with a GitHub account. The free plan included 2,000 code completions and 50 chat messages per month, with access to Claude 3.5 Sonnet and GPT-4o. No credit card was required, and the free SKU worked in VS Code, Visual Studio, JetBrains IDEs, and on GitHub.com. The move was widely seen as a response to growing competition from free tools like Codeium, Cline, and the rapid rise of Cursor, and as a strategic decision to lower the barrier to entry for the 150 million developers on GitHub at the time.
On February 6, 2025, GitHub announced agent mode for Copilot in VS Code along with Next Edit Suggestions, a feature that predicts the developer's next likely edit based on the history of changes. Agent mode could autonomously plan and execute multi-step tasks: reading files, proposing edits across multiple files, running terminal commands, monitoring test output, and auto-correcting errors in a loop until the task was complete. Agent mode also added support for the Model Context Protocol (MCP), letting users connect Copilot to external data sources and tools. The same announcement included the general availability of Copilot Edits.
In April 2025, agent mode reached general availability across VS Code. On May 17, 2025, GitHub announced the coding agent, a more autonomous mode that operates asynchronously in the cloud. Developers could assign a GitHub issue to Copilot (by tagging @copilot), and the coding agent would spin up a secure cloud development environment powered by GitHub Actions, write the code, run tests, perform self-review using Copilot code review, and open a draft pull request with commits pushed in real time. At Microsoft Build on May 19, 2025, GitHub also announced that Copilot Chat in VS Code would be open sourced and that agent mode would expand to JetBrains, Eclipse, and Xcode.
The coding agent was made generally available in the second half of 2025. By July 2025, GitHub Copilot surpassed 20 million total users, a 5 million increase in just three months. The Copilot Workspace technical preview was sunset on May 30, 2025, with GitHub stating that the architectural lessons (sub-agent design, issue to pull request workflow, asynchronous execution) had been rebuilt into the production coding agent. In September 2025, GitHub deprecated the earlier Copilot Extensions framework (based on GitHub Apps) in favour of MCP, reflecting a broader industry shift toward standardised AI tool integration.
When GitHub Copilot launched in 2021, it was powered exclusively by OpenAI Codex, a model fine tuned from GPT-3 on a large dataset of publicly available source code. Codex could interpret natural language prompts and translate them into code, and it supported over a dozen programming languages, with strongest performance on Python, JavaScript, TypeScript, Ruby, and Go. The Codex API itself was deprecated by OpenAI in March 2023, by which point Copilot had moved its chat features to GPT-4 while keeping a faster, smaller model for inline completions.
The single model architecture began to shift in 2023. Copilot Chat shipped on GPT-3.5 initially and was upgraded to GPT-4 in November 2023 as part of the broader Copilot X rollout. For inline completions, GitHub used a custom, smaller model optimised for low latency. A GPT-4o based completion model was rolled out across most of 2024 to improve both quality and latency.
The biggest change came at GitHub Universe 2024, when Copilot officially became multi-model. As of early 2026, developers can choose between models from four different providers depending on their plan.
| Model | Provider | Notes |
|---|---|---|
| GPT-4.1 | OpenAI | Default model for chat, agent mode, and completions |
| GPT-4o | OpenAI | Fast, multimodal model |
| GPT-5 (preview) | OpenAI | Advanced reasoning, available in preview |
| o1 / o3 (preview) | OpenAI | Reasoning-focused models |
| Claude 3.5 Sonnet | Anthropic | Strong coding performance, added Oct 29, 2024 |
| Claude 3.7 Sonnet | Anthropic | Extended thinking capabilities |
| Claude Opus 4 | Anthropic | Enterprise tier only |
| Gemini 1.5 Pro | Initial Gemini integration, late 2024 | |
| Gemini 2.5 Pro | Hosted on Google Cloud | |
| Gemini 3 Flash | Low-latency option | |
| Gemini 3 Pro | Full-featured Google model | |
| Grok Code Fast | xAI | Added in 2025 |
For inline code completions (the core autocomplete feature), GitHub continues to use a smaller, custom model optimised for low latency. The larger models in the table above are primarily used for Copilot Chat, agent mode, and the coding agent. The choice of which model is appropriate is left to the developer through a model picker in the IDE, and organisations on Business and Enterprise plans can restrict which models employees can select for policy reasons.
GitHub Copilot supports code generation in virtually any programming language present in its training data, but performs best with:
| Language | Support level |
|---|---|
| Python | Strong |
| JavaScript | Strong |
| TypeScript | Strong |
| Go | Strong |
| Ruby | Strong |
| Java | Strong |
| C / C++ | Strong |
| C# | Strong |
| Rust | Good |
| PHP | Good |
| Swift | Good |
| Kotlin | Good |
| Shell / Bash | Good |
| SQL | Good |
Quality varies significantly across languages with smaller training corpora. Niche or proprietary languages tend to produce hallucinated APIs and incorrect syntax more often than the mainstream choices in the table.
The original and most fundamental feature of GitHub Copilot is inline code completion. As a developer types, Copilot analyses the context and offers suggestions that can range from completing a single line to generating entire functions, classes, or code blocks. Developers can accept suggestions with Tab, cycle through alternatives, or dismiss them. Copilot adapts to the coding style and patterns in the current file and project. It can infer intent from comments written in natural language and generate corresponding code implementations.
The context window for completions is constructed from the code immediately around the cursor, related open files, file names, and import statements. The system uses heuristics to decide what to include before sending the prompt to the model. Earlier versions had a relatively small context window, which limited the quality of suggestions on large files; later versions of the GPT-4o completion model widened the window substantially.
Introduced in February 2025, Next Edit Suggestions (NES) represents an evolution beyond traditional completions. Rather than simply completing the current line, NES predicts what the developer will want to edit next based on their recent changes. For example, if a developer renames a variable, NES will suggest updating all other references to that variable throughout the file. If a developer adds an argument to a function signature, NES will propose corresponding edits at each call site. NES uses a custom model trained specifically for this purpose, incorporating edit history rather than just the current state of the code, and is available in VS Code, Visual Studio, JetBrains IDEs, Xcode, and Eclipse.
Copilot Chat is a conversational interface embedded directly within supported IDEs and on GitHub.com. Developers can ask questions about their code, request explanations of complex logic, get help with debugging, and ask Copilot to generate or refactor code. Chat is context aware, meaning it can reference the currently open file, selected code, terminal output, and error messages. Chat supports multiple models and can be used in several modes:
Chat also exposes slash commands for common operations such as /explain, /fix, /tests, /doc, and /new (which scaffolds a new project), as well as participants prefixed with @ that route the conversation to a specialised handler, such as @workspace for project-wide questions or @terminal for shell commands. Inline Chat opens a focused chat window directly over selected code, which is useful for quick refactors.
Copilot Edits, previewed in November 2024 and made generally available in February 2025, lets a developer describe a change in natural language and have Copilot apply it across a defined working set of files. The interface includes a panel where files can be dragged in, a chat input for describing the desired change, and an inline diff view for accepting or rejecting individual hunks. Edits is built on top of the same chat models and works well for cross-cutting refactors, schema changes, and reformatting tasks that span more than one file.
Agent mode, announced in February 2025, transforms Copilot from a suggestion tool into an autonomous coding agent. When given a task in natural language, agent mode:
Agent mode supports Model Context Protocol (MCP) integration, allowing it to connect to external data sources and tools. Users can configure MCP servers in a JSON file or through the VS Code UI. Common MCP servers include connections to databases, internal APIs, observability platforms, and ticketing systems.
The coding agent, announced in May 2025 and made generally available later in 2025, goes further than agent mode by operating asynchronously in the cloud. Developers can assign a GitHub issue to Copilot (by tagging @copilot or assigning the issue through the GitHub UI), and the coding agent will:
The coding agent can be extended with MCP servers to access external databases, APIs, and development tools, and runs in a sandboxed environment with network access restricted by default. Organisations can configure which repositories the agent has access to and what runners it uses, similar to the configuration model for GitHub Actions.
Copilot code review allows developers to request an AI review of their pull requests directly on GitHub.com. Copilot analyses the diff, identifies potential bugs, security issues, and style inconsistencies, and leaves review comments with suggested fixes. Organisations can configure automatic Copilot reviews for all pull requests, including those from contributors without a Copilot license. As of 2025, Copilot code review is available in VS Code, Visual Studio, JetBrains IDEs, and via the GitHub CLI. The coding agent uses code review internally as a self-check step before marking a draft pull request as ready for human review.
Copilot for CLI brings AI assistance to the command line. Developers can ask Copilot to explain commands, compose complex shell operations, and troubleshoot errors. The CLI supports natural language queries and returns executable commands with explanations. Copilot CLI also supports MCP server integrations for custom tooling and plugins.
Copilot Workspace, previewed on April 29, 2024, was an experimental browser-based development environment. Given a GitHub issue written in natural language, Copilot Workspace could generate a specification, propose a plan, and produce actual code changes, all within a browser interface. The workflow followed four stages: brainstorm (clarifying requirements), plan (a step-by-step implementation plan), implement (writing and testing the code), and verify (running tests). Workspace used multiple specialised AI agents for different stages, including a brainstorm agent, a plan agent, and an implementation agent.
The technical preview was sunset on May 30, 2025. GitHub took the architectural lessons learned (sub-agent design, issue-to-pull-request workflow, asynchronous execution) and rebuilt them into the Copilot coding agent, which reached general availability in the second half of 2025.
Copilot Extensions, announced in limited public beta on May 21, 2024, let third-party developers build chat participants and skills that could be invoked from Copilot Chat. Launch partners included DataStax, Docker, LambdaTest, LaunchDarkly, McKinsey & Company, Microsoft Azure and Teams, MongoDB, Octopus Deploy, Pangea, Pinecone, Product Science, ReadMe, Sentry, and Stripe. In September 2025, GitHub deprecated the GitHub Apps based Extensions framework in favour of MCP servers, on the basis that MCP had become an industry standard.
GitHub Spark, previewed at GitHub Universe 2024, is a natural language micro-app builder. Users describe an app in plain English and Spark scaffolds a small web app, hosts it on GitHub Pages with a GitHub repository underneath, runs serverless functions through GitHub Actions, and stores data in Azure Cosmos DB. Spark targets fun and personal software rather than enterprise productivity apps and is part of GitHub's broader vision of bringing software creation to a wider audience.
GitHub Copilot is available across a wide range of development environments. Feature availability varies by IDE.
| IDE | Code completion | Chat | Agent mode | Code review | NES |
|---|---|---|---|---|---|
| Visual Studio Code | Yes | Yes | Yes | Yes | Yes |
| Visual Studio | Yes | Yes | Yes | Yes | Yes |
| JetBrains IDEs | Yes | Yes | Yes | Yes | Preview |
| Neovim / Vim | Yes | Yes | No | No | No |
| Xcode | Yes | Yes | No | No | Preview |
| Eclipse | Yes | Yes | No | No | Preview |
| Azure Data Studio | Yes | No | No | No | No |
| GitHub.com | N/A | Yes | N/A | Yes | N/A |
VS Code receives the most complete feature support, as both VS Code and Copilot are Microsoft products with tightly integrated development teams. Agent mode rolled out to JetBrains, Eclipse, and Xcode through 2025 following its initial VS Code launch.
GitHub Copilot is available across five plan tiers as of early 2026.
| Plan | Price | Code completions | Chat messages | Premium requests | Coding agent | Code review | Key features |
|---|---|---|---|---|---|---|---|
| Free | $0 | 2,000 per month | 50 per month | None | No | No | Basic completions and chat; Claude 3.5 Sonnet and GPT-4o |
| Pro | $10 per month | Unlimited | Unlimited | 300 per month | Yes | Yes | Multiple model choices; unlimited suggestions |
| Pro+ | $39 per month | Unlimited | Unlimited | 1,500 per month | Yes | Yes | All Pro features plus expanded premium requests and GitHub Spark |
| Business | $19 per user per month | Unlimited | Unlimited | 300 per user per month | Yes | Yes | Organisational management; usage metrics; IP indemnity; data excluded from training |
| Enterprise | $39 per user per month | Unlimited | Unlimited | 1,000 per user per month | Yes | Yes | All Business features; knowledge bases; advanced models (Claude Opus); MCP server integration |
Premium requests are consumed when using advanced models (beyond the default) or when using features like the coding agent. Verified students, educators, and maintainers of popular open-source projects receive Copilot Pro for free.
GitHub has also experimented with usage-based billing for premium model requests above the included quota. Customers who exceed their plan's premium request allowance can be charged on a per-request basis, with rates that vary by model. The pricing model has evolved several times since 2022; the initial $10 per month flat fee for individuals has been retained, but the introduction of premium request quotas, the free tier, and usage-based overage charges have made the actual cost of heavy Copilot usage significantly less predictable than it once was.
GitHub Copilot has seen rapid adoption since its launch, and the platform's growth has consistently exceeded internal projections. Microsoft reports user and subscriber counts each quarter on its earnings calls, providing a fairly clear public picture of growth.
| Metric | Value | Date |
|---|---|---|
| Paid subscribers | 1 million+ | February 2023 (Q2 FY23 earnings) |
| Paid subscribers | 1.3 million | April 2024 (Q3 FY24 earnings) |
| Paid subscribers | 1.8 million | October 2024 (Q1 FY25 earnings) |
| Enterprise customers | 77,000+ | January 2025 (Q2 FY25 earnings) |
| Total users | 20 million+ | July 2025 |
| Paid subscribers | 4.7 million | January 2026 |
| Fortune 100 adoption | ~90% | 2025 |
| Annual recurring revenue | $2 billion+ | 2025 |
| Code written by Copilot (active users) | ~46% | 2025 |
| AI coding tool market share | ~42% | 2025 |
In early 2024, Microsoft CEO Satya Nadella stated on an earnings call that GitHub Copilot had become a larger business than all of GitHub was at the time Microsoft acquired GitHub in 2018 for $7.5 billion. Copilot accounted for over 40% of GitHub's revenue growth during that period. By late 2025, GitHub recorded approximately 75% year-over-year growth in Copilot paid seats, and Copilot adoption was a primary driver of overall GitHub revenue, which by then was approaching $2 billion in annual run rate.
Adoption inside large enterprises has been particularly notable. GitHub reports that approximately 90% of the Fortune 100 uses Copilot at some level, with major customers including Accenture, Coca-Cola, Goldman Sachs, BBVA, and the U.S. federal government. Microsoft has used Copilot as a flagship reference for its broader AI sales push, and many of the Microsoft 365 Copilot deals announced in 2024 and 2025 included GitHub Copilot as part of the bundle.
Multiple research studies have examined GitHub Copilot's effect on developer productivity, with results that are generally positive but more nuanced than the marketing material suggests.
The most widely cited study is Peng, Kalliamvakou, Cihon, and Demirer's 2023 paper "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot" (arXiv:2302.06590), published by researchers at GitHub Next and the Microsoft Office of the Chief Economist. The experiment ran from May 15, 2022 to June 20, 2022, just before Copilot's general availability launch, and used a randomised controlled trial design with 95 professional developers. Each participant was asked to implement an HTTP server in JavaScript; half were given access to Copilot and half were not. Developers using Copilot completed the task in an average of 1 hour and 11 minutes, compared with 2 hours and 41 minutes for the control group. The 55.8% speedup was statistically significant (p = 0.0017), with a 95% confidence interval of [21%, 89%].
The paper also reported that developers with less programming experience, older programmers, and developers who spent more hours per day coding benefited most from Copilot. The implicit explanation was that Copilot reduces the cognitive cost of recall on routine boilerplate, which disproportionately helps developers who would otherwise spend more time looking things up.
A subsequent field experiment by Cui, Demirer, Jaffe, Musolff, Peng, and Salz, conducted across three companies (Microsoft, Accenture, and an unnamed Fortune 100 firm), found a 26.08% increase in the number of tasks completed by developers using Copilot, with the largest productivity gains observed among less experienced developers. GitHub's own internal research has reported that between 60% and 75% of Copilot users feel more fulfilled with their work and less frustrated when coding. 73% of users in those surveys said Copilot helped them stay in a state of flow, and 87% said it reduced mental effort during repetitive coding tasks. These figures are commonly quoted in GitHub marketing material and have also been cited in academic papers.
Not all research has been uniformly positive. A study by Uplevel Data Labs in 2024 found that while developers reported subjective productivity improvements, Copilot users had a significantly higher bug rate compared with non-users, and raw throughput (measured by pull requests) remained roughly constant. The implication was that Copilot may sometimes trade code quality for speed.
GitClear, a code review analytics company, published a series of widely discussed reports in 2024 and 2025 based on roughly 153 million changed lines of code authored between January 2020 and December 2023. The 2024 report (Coding on Copilot: 2023 Data Suggests Downward Pressure on Code Quality) reported that code churn, defined as the percentage of lines reverted or updated less than two weeks after being authored, was projected to double in 2024 compared with the 2021 pre-AI baseline. The percentage of copy/pasted code rose from 8.3% to 12.3% between 2021 and 2024, while lines classified as refactoring fell from 25% to under 10%. The 2025 follow up (AI Copilot Code Quality: 2025 Data Suggests 4x Growth in Code Clones) reported continued increases in code cloning and concluded that AI assistants were correlated with lower code reuse.
A Microsoft internal study from 2024 similarly found limited measurable telemetry impact on throughput, though developers consistently reported perceived productivity gains, particularly for boilerplate and repetitive code. A more sceptical 2025 study from METR, which focused on senior open source developers using Cursor rather than Copilot, found that AI assistance actually slowed the studied developers down by approximately 19%, even though those same developers believed they had been faster. The METR result attracted significant attention because it ran against both the prevailing narrative and the developers' own intuitions, and it has frequently been cited in discussions of Copilot as well.
The collective picture from research is that AI coding assistants speed up well defined boilerplate and lookup tasks substantially, that they are perceived as more helpful than measured throughput would suggest, and that they may trade some code quality and architectural rigour for short term speed. The size of the effect depends heavily on the task, the developer's experience level, and the codebase.
On November 3, 2022, attorneys Joseph Saveri and Matthew Butterick filed a class action lawsuit in the U.S. District Court for the Northern District of California against GitHub, Inc., Microsoft Corporation, and a set of OpenAI entities (case 4:22-cv-06823-JST, Judge Jon Tigar presiding). The case, Doe 1 et al. v. GitHub, Inc. et al., was filed on behalf of two anonymous plaintiffs and a proposed class of open source programmers who alleged that Copilot's training data, sourced from public GitHub repositories, violated the licensing terms attached to that code.
The core legal argument was that Codex and Copilot were trained on billions of lines of open source code without complying with the requirements of licenses such as the GPL, MIT License, and Apache License. These licenses typically require attribution, inclusion of the license text, or other conditions that Copilot did not fulfil when reproducing or paraphrasing code. The original complaint included 22 claims, ranging from violation of the Digital Millennium Copyright Act (DMCA) sections 1202(a) and 1202(b) (removal or alteration of copyright management information), breach of contract under the GitHub Terms of Service, tortious interference, fraud, false designation of origin, unjust enrichment, unfair competition, breach of GitHub's Privacy Policy, violation of the California Consumer Privacy Act, and negligence.
In May 2023, Judge Tigar dismissed several of the original claims with leave to amend, including the tortious interference, fraud, false designation of origin, unjust enrichment, unfair competition, privacy policy, CCPA, and negligence claims. The DMCA claims under section 1202(b) were initially allowed to proceed, but only on a narrow reading.
In January 2024, the court dismissed the majority of the remaining original claims, including most DMCA allegations. Judge Tigar ruled that plaintiffs had failed to show that Copilot distributed code without copyright management information. However, the court allowed certain claims under DMCA section 1202(a)(1) to proceed.
On July 5, 2024, Judge Tigar dismissed the remaining DMCA section 1202(b) claims. The court ruled that DMCA section 1202(b) imposes an identicality requirement: liability attaches only when copyright management information is removed or altered from an identical copy of a copyrighted work. Because Copilot's outputs were typically paraphrased or transformed rather than identical reproductions, the court held that section 1202(b) did not apply. This left only a narrow set of breach of contract claims against GitHub.
On October 7, 2024, the plaintiffs filed a petition for permission to appeal to the U.S. Court of Appeals for the Ninth Circuit. The petition presents what plaintiffs described as an unresolved question of statutory interpretation under DMCA section 1202(b): whether section 1202(b) requires that copies of works be identical to the original for liability to attach. The Ninth Circuit assigned the case docket number 24-7700, and the appeal was being briefed through 2025. A ruling in favour of the plaintiffs would re-open a wide range of generative AI training cases, while a ruling in favour of the defendants would effectively settle the DMCA identicality question in the AI context for the Ninth Circuit.
The broader open source community and several civil society organisations filed amicus briefs on both sides of the appeal, reflecting the precedential significance of the case for both AI training and software copyright generally.
In response to legal concerns, GitHub introduced IP indemnity for Copilot Business and Enterprise customers in late 2023. Under this policy, Microsoft assumes legal liability if Copilot generates code that results in an intellectual property infringement claim, provided the customer is on a qualifying plan and has enabled GitHub's duplicate code filtering feature. The filter is designed to suppress suggestions that closely match public code in the training data. The indemnity is similar to indemnities Microsoft offers for its Microsoft 365 Copilot products and is widely seen as a sales tool for the Business and Enterprise tiers, allowing customers to shift legal risk to the vendor.
The AI coding assistant market has grown rapidly since Copilot's launch, with a small number of well funded competitors offering meaningfully different approaches. By late 2025, Copilot still held roughly 42% of the paid market by revenue, but competition was intensifying on multiple fronts.
| Tool | Developer | Pricing (individual) | Key differentiator |
|---|---|---|---|
| GitHub Copilot | GitHub / Microsoft | Free to $39 per month | Largest user base; deep GitHub integration; multi-model support |
| Cursor | Anysphere | $20 per month | Fork of VS Code with AI-native editor; strong agent capabilities |
| Claude Code | Anthropic | Usage-based | Command-line tool; autonomous multi-step coding; deep codebase understanding |
| Cline | Cline community | Free / open source | VS Code extension agent that works with any model API |
| Aider | Aider community | Free / open source | Command-line pair programmer with git integration |
| Windsurf | Codeium | $15 per month | AI-native IDE; Cascade agentic flow |
| Codeium | Exafunction | Free (unlimited) | Free tier with unlimited completions; strong privacy guarantees |
| Amazon Q Developer | Amazon | Free to $19 per month | Deep AWS integration; code transformation for Java upgrades |
| Tabnine | Tabnine | Free to $12 per month | On-premise deployment; privacy-first approach |
| Gemini Code Assist | Free to $19 per user per month | Google Cloud integration; Gemini model access | |
| JetBrains AI / Junie | JetBrains | Subscription bundled | Native integration with JetBrains IDEs |
Cursor emerged as Copilot's most visible competitor in 2024 and 2025. Built by Anysphere, a startup founded in 2022 by Michael Truell, Sualeh Asif, Arvid Lunnemark, and Aman Sanger while they were MIT students, Cursor is a fork of VS Code that has been redesigned around an AI-native workflow with built-in chat, multi-file edits, and an agent. The OpenAI Startup Fund led an $8 million seed in October 2023, with Nat Friedman (former GitHub CEO) and Dropbox co-founder Arash Ferdowsi among the angels. Anysphere raised $100 million in December 2024 at a roughly $2.4 billion valuation, and by November 2025 was in talks to raise approximately $2.3 billion at a $29.3 billion valuation.
Revenue growth at Cursor has been extraordinary. The company hit approximately $100 million in annualised revenue in January 2025, $500 million by June 2025, and crossed $1 billion later in 2025. Industry reporting has called the trajectory from zero to $2 billion ARR in roughly three years the fastest scaling B2B software ramp on record. The product is widely regarded as superior to Copilot for agentic, multi-file workflows, while Copilot still has an edge in enterprise distribution, free tier reach, and deep GitHub integration.
Claude Code, released by Anthropic in 2025, took a different approach as a terminal-based agent rather than an IDE plugin. By late 2025, Claude Code had surged in developer awareness, reaching roughly 57% awareness in surveys by January 2026 with around 18% active workplace usage. Its underlying capability, autonomous multi-step code generation, is converging with Copilot's coding agent, although the workflow position (terminal vs IDE) is different.
Cline, an open-source VS Code extension, popularised the pattern of a model agnostic agent that could read and edit files, run terminal commands, and iterate on results, all driven by whichever model API the user chose to provide. Cline's permissive licence and BYO key model made it a popular alternative for developers who wanted full control over which model they used and how much they spent. Aider, a command-line pair programmer that emphasises git integration, occupied a similar niche.
Other meaningful competitors include Windsurf (Codeium's AI-native IDE), Codeium itself with its free, unlimited tier, Amazon Q Developer (deeply integrated with AWS and notable for its Java code transformation feature), Tabnine (focused on on-premise deployment and privacy), Gemini Code Assist for Google Cloud users, and JetBrains's own AI Assistant and Junie agent, which are bundled with JetBrains IDEs.
Despite the increasing competition, Copilot retained its first-mover advantage in 2025 through the sheer scale of its installed base, the depth of its integration with GitHub, and Microsoft's enterprise sales motion. The strategic shift toward multi-model support was widely read as an admission that no single model would win the coding category and that the path to long term dominance was through distribution and integration rather than model exclusivity.
GitHub Copilot has generated three distinct categories of criticism since its launch: licensing and copyright, security, and code quality.
The Doe v. GitHub lawsuit, discussed in detail above, is the most visible manifestation of the licensing concerns. The underlying complaint, that training large language models on open source code without complying with attribution or share-alike provisions of the source licenses, is widely shared in the open source community even among developers who are not parties to the case. The Free Software Foundation, Software Freedom Conservancy, and other organisations have published critiques of Copilot's training practices since 2021, and several prominent maintainers (including Drew DeVault and Tim Davis) have written publicly about Copilot reproducing their code verbatim without attribution.
GitHub's response has been a mix of technical and contractual measures. The duplicate code filter, available to all paid subscribers, suppresses suggestions that match public code beyond a configurable similarity threshold. The IP indemnity for Business and Enterprise customers shifts legal risk to Microsoft. And GitHub has argued, in court and in public statements, that training on publicly available code is a fair use under U.S. copyright law and that licence terms generally do not extend to training of statistical models. The legal merits of these positions are unsettled as of 2025 and depend in part on the Ninth Circuit's eventual ruling in Doe v. GitHub.
Researchers have identified several security-related concerns with AI-generated code from Copilot. The earliest and most influential study was Pearce, Ahmad, Tan, Dolan-Gavitt, and Karri's 2021 NYU paper "Asleep at the Keyboard? Assessing the Security of GitHub Copilot's Code Contributions" (arXiv:2108.09293). The researchers constructed 89 scenarios designed to elicit code with potential vulnerabilities and had Copilot generate 1,689 programs across them. They found that approximately 40% of the generated programs contained vulnerabilities, spanning a range of Common Weakness Enumeration (CWE) categories. The paper won a Distinguished Paper Award at the IEEE Symposium on Security and Privacy and remains the canonical reference for security concerns about AI coding assistants.
A follow up empirical study analysing Copilot-generated code snippets on GitHub found that 29.5% of Python snippets and 24.2% of JavaScript snippets contained security weaknesses. These spanned 43 Common Weakness Enumeration categories, including CWE-330 (use of insufficiently random values), CWE-94 (improper control of code generation, i.e. code injection), and CWE-79 (improper neutralisation of input, i.e. cross-site scripting). A 2025 study evaluating Copilot's own code review feature found that it frequently failed to detect critical vulnerabilities like SQL injection, cross-site scripting, and insecure deserialisation. Instead, the automated reviews tended to focus on low-severity issues such as coding style and typographical errors.
A controlled user study, often cited as Perry et al., found that developers using Copilot were more likely to submit insecure code compared with developers coding without AI assistance, and that these developers also expressed higher confidence in the security of their submissions despite the presence of vulnerabilities. The combination, more bugs and more confidence, is particularly concerning from a security standpoint because overconfidence reduces the likelihood that vulnerabilities will be caught during review.
An analysis of roughly 20,000 repositories with active Copilot usage found that over 1,200 (about 6.4%) had leaked at least one secret, such as an API key or credential. This rate was about 40% higher than the baseline across all public repositories, likely because Copilot can suggest code patterns that inadvertently expose sensitive data, such as inlined credentials in example configuration files. GitHub has responded with secret scanning features and by encouraging organisations to use Copilot's content exclusion settings to prevent sensitive files from being sent to the model.
The GitClear studies described in the productivity research section are the most cited evidence that Copilot may degrade code quality over time. The 2024 report's central finding, that code churn (the percentage of lines reverted within two weeks) was projected to double in 2024 compared with 2021, has been cited extensively. The 2025 follow up reported a roughly 4x increase in code cloning, suggesting that AI assistants encourage developers to duplicate logic rather than refactor.
A related concern is the hallucination of APIs. Copilot has occasionally suggested function calls or import statements for libraries that do not exist or for functions that the library does not expose. This is particularly common in lower resource languages or in libraries that have recently changed their API surface. Researchers have argued that hallucinated APIs in suggested code create a category of supply chain risk: an attacker who notices a hallucinated package name can register a malicious package under that name (slopsquatting) and rely on developers' acceptance of Copilot suggestions to install it.
A recurring privacy concern is that code sent to Copilot for completion or chat is transmitted off the user's machine to a hosted model. For individuals, this is generally acceptable, but for organisations with sensitive intellectual property, it can be a barrier to adoption. GitHub has addressed this in several ways: the Business and Enterprise plans contractually exclude customer code from training data, content exclusions let administrators specify file patterns that should not be sent to the model, and an upcoming on-premise variant of Copilot has been signposted for customers with the strictest data residency requirements.