GitHub has announced the general availability of its AI pair programmer Copilot for Business, boasting an improved AI model and new features designed to ensure code safety.
Businesses are now able to sign up for the subscription-based system, which suggests code and functions to developers within their chosen text editor.
The business version completed a brief beta test in December and offers organisation-wide policies, plus the newest Copilot model which can automatically block coding patterns containing security flaws.
This allows it to detect common vulnerabilities such as SQL injections and hard-coded credentials, even in incomplete code. This can empower developers to address flaws before they’re present in code, in contrast to traditional static analysis tools which assess a completed software build.
Subscribers to the new business tier will also have exclusive support for virtual private networks (VPNs) within the system, even with self-signed certificates, to allow flexible use of the system across any organisation’s developer environment.
Copilot for Business costs $19 (£16) per month, per user. It offers companies licence management options, with administrators able to decide which teams and developers receive licenses for the tool.
“In the coming years, we will integrate AI into every aspect of the developer experience - from coding to the pull request to code deployments - so developers can build their best in a world where all organisations will be more dependent on their success than ever,” wrote Thomas Dohmke, CEO of GitHub in a blog post.
“GitHub Copilot for Business is the first stride in this future, a future that will push the boundaries for all developers.”
The idea for the business version of Copilot was driven by user feedback. Developers reported wide-scale interest in using the tool professionally but since it was only able to be purchased by an individual, the logistics of getting every developer within an organisation their own license was tiring.
Speaking to IT Pro at GitHub Universe in November 2022, Jeremy Epling, VP of product at GitHub, said the move was expected to drive the adoption of Copilot within businesses.
GitHub made Copilot for individuals generally available in June 2022, and Copilot for Business available as a beta product in December.
When Copilot for individuals first became generally available, it was already being used by developers to generate 27% of their code. At the time of writing, this has jumped to 46% of code across all languages and as high as 61% for developers using Java.
Internal research found that 90% of developers who used Copilot completed tasks faster, and 75% staved off burnout by using the tool to feel more focused and fulfilled.
Six steps to machine learning success
The path toward leveraging the full power of machine learning
Copilot is a generative AI model which utilises Codex, a natural language processing system made by OpenAI, which claims to be proficient in a number of the top programming languages.
While Copilot has been trained on all languages that appear in public GitHub repositories, it’s better at languages that are used frequently.
GitHub has stated in the past that it’s especially accomplished at producing JavaScript, while Microsoft guidance on Codex lists Python as its speciality.
GitHub has stressed that businesses' code is not stored on its end, stating "we won’t retain code snippets, store or share your code regardless if the data is from public repositories, private repositories, non-GitHub repositories, or local files".
Alongside the release, GitHub has updated both tiers of Copilot with AI improvements such as a paradigm by which Copilot makes code suggestions called Fill-In-the-Middle (FIM).
This considers both the prefix of written code as well as code suffixes included in its model, leaving a gap’ in the middle of the code which Copilot then fills. The firm stated that this has improved the quality of code suggestions, without adding latency to the system.
GitHub has also added a new lightweight Copilot extension for Visual Studio code, which leverages user context such as how many of its suggestions have been accepted to prevent unnecessarily distracting developers with unwanted prompts.
