Coding Tools Go Plug-and-Play

Visual Studio Code now lets GitHub Copilot Business and Enterprise customers plug in their own language-model keys, turning the editor into a front end for outside artificial intelligence systems. (github.blog) GitHub posted the change on April 22, 2026, and Microsoft shipped it in Visual Studio Code 1.117 the same day. The feature works in VS Code Chat, including the built-in plan agent and custom agents, but GitHub said it does not apply to code completions. (github.blog) (code.visualstudio.com) Teams can connect API keys from providers including Anthropic, Gemini, OpenAI, OpenRouter, Azure, Ollama, and Foundry Local, then keep working inside the same Copilot workflow. GitHub said usage is billed by the outside provider and does not count against GitHub Copilot request quotas. (github.blog) (code.visualstudio.com) That setup shifts Copilot from a fixed assistant into a routing layer: the chat box stays the same while the model behind it can change for cost, compliance, or performance reasons. Microsoft said administrators can enable or disable the policy in GitHub.com settings and control which providers their organization can use. (code.visualstudio.com) (github.blog) OpenAI made a parallel move on April 21, 2026, with ChatGPT Images 2.0, a new image model aimed at work that depends on readable text, charts, diagrams, and multilingual layouts. OpenAI said the model improves text rendering, multilingual support, and visual reasoning. (openai.com) (macrumors.com) OpenAI’s pitch is less about surreal art than office output: slides, infographics, user-interface mockups, and other images where garbled labels used to give away machine-made work. MacRumors reported the new model was presented as better at following detailed instructions and producing more usable professional graphics. (openai.com) (macrumors.com) The common thread is that AI tools are being folded into ordinary software jobs instead of sitting beside them as separate demos. OpenAI said this week that Codex had reached 4 million weekly active users and that consulting firms including Accenture, Capgemini, CGI, Cognizant, Infosys, PwC, and Tata Consultancy Services would help companies deploy it across the software development lifecycle. (openai.com) That changes what “using AI at work” looks like in practice. Microsoft’s own GitHub Copilot training materials tell developers to test and validate an agent’s output before merging, and GitHub said last year that its coding agent had started automatically running security and quality analysis on code it generated. (learn.microsoft.com) (github.blog) Hiring is likely to follow the workflow. OpenAI says one of its Codex deployment roles is for people who are “active power” users of AI coding tools with a view on what makes engineers more productive, which points to employers valuing judgment over raw autocomplete speed. (openai.com) The new baseline is not “can you prompt a model,” but whether you can steer one, check its work, and ship code that another human can maintain. VS Code’s new key-switching option makes that expectation easier to enforce inside the tools many teams already use. (github.blog) (learn.microsoft.com)