Microsoft has introduced a new lightweight coding model called MAI-Code-1-Flash, designed to assist developers in fast and iterative coding tasks, such as writing and refining code in real time. This model is part of GitHub Copilot, an AI tool that helps developers write code more efficiently. Since its launch at the Build conference, MAI-Code-1-Flash has been used in production alongside other main coding models, allowing Microsoft to observe its real-world performance in developer workflows. The results show that when combined with VS Code, a popular code editor, the model provides high-quality code while reducing the use of tokens—units of data processed by AI models—which helps developers stay within their usage limits. When evaluating coding models, developers and companies consider both the quality of the generated code and the efficiency of token usage. High-quality code is essential, but reducing token consumption helps lower costs. Microsoft studied how well MAI-Code-1-Flash balances these two factors by analyzing data from developers who used the model in VS Code Copilot Chat. They looked at how often developers accepted the AI’s suggestions, how long the code remained useful after editing, and how many tokens were used per task. The findings suggest that MAI-Code-1-Flash performs well in both areas, offering better code quality than some similar models while using fewer tokens. Compared to models like Claude Haiku 4.5 and GPT-5.4 Mini, MAI-Code-1-Flash shows higher quality in its outputs. Larger models, such as GPT-5.6 Luna and Kimi K2.7 Code, provide even better quality but require significantly more tokens per interaction, which increases costs. Microsoft’s data also showed that developers using MAI-Code-1-Flash were more likely to continue using it compared to users of other models. Specifically, users were 6% more likely to return within two days than those using GPT-5.4 Mini and 11% more likely than those using Claude Haiku 4.5. Despite this higher engagement, token usage was 13% lower than with GPT-5.4 Mini and 11% lower than with Claude Haiku 4.5. To further test the model's effectiveness, Microsoft conducted A/B tests in VS Code Copilot Chat, where the model was automatically assigned to users without them choosing it. This allowed researchers to assess how engaging and useful the model was, independent of user preferences or specific tasks. The results supported the earlier findings: MAI-Code-1-Flash was more engaging and used tokens more efficiently. It is designed for lightweight, iterative tasks like exploring code, making small changes, writing tests, and fixing bugs. Its ability to provide concise answers when needed and expand on complex requests when necessary makes it both fast and efficient. MAI-Code-1-Flash is now available in GitHub Copilot, and developers can test it alongside their preferred models to see how it fits into their workflows. User feedback will help shape future improvements. While MAI-Code-1-Flash is a step forward in lightweight coding models, Microsoft plans to release more advanced versions later this year, capable of handling more complex tasks with deeper reasoning and higher quality. These models will be trained using clean, traceable data and will continue to improve through data gathered from public usage, without using data from Copilot Business or Copilot Enterprise for training.