3 min Devops

Microsoft MAI Code 2 generates code at a fraction of the price

Microsoft MAI Code 2 generates code at a fraction of the price

Microsoft is introducing the MAI Code 2 and MAI Code 2 Flash programming models. According to Microsoft, MAI Code 2 comes very close to GPT-6.1 Sol and Claude Opus 5.5, at a fraction of the cost. Both models are available in GitHub Copilot.

MAI Code 2 focuses on the work developers do every day: deploying code to production, thorough debugging, multi-file implementations, and reasoning across an entire repository. It also covers UI work. The smaller variant, MAI Code 2 Flash, is intended for large volumes of simpler tasks, such as code explanations, targeted adjustments, routine fixes, and rapid iterations.

Microsoft states that the models are “not yet tuned for harder problems,” such as fundamental scientific research and agentic work with a very long time horizon. The price-performance ratio is the top priority.

Tip: Microsoft is enhancing AI agents with new models

98 percent of GPT-6.1 Sol for 40 percent of the price

Microsoft measured the results across four benchmarks aligned with tasks it sees in GitHub Copilot: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1. On these benchmarks, MAI Code 2 achieves 98 percent of GPT-6.1 Sol’s performance at 40 percent of the cost. Compared to Claude Opus 5.5, it achieves 96 percent of the performance at just 5 percent of the cost. These figures come from Microsoft’s internal evaluation platform.

Visual work also receives attention. MAI Code 2 is optimized for dashboards and visualizations. In blind evaluations of front-end apps, evaluators preferred the results from MAI Code 2 over those from GPT-6 Sol in 64 percent of cases.

According to Microsoft, the Flash variant performs comparably to GPT-6 Luna and Claude Haiku 5.5 on three benchmarks for routine coding tasks, at 65 and 20 percent of the cost, respectively.

Based on Code 2 Flash, Microsoft optimized a separate model, MAI Work Flash, for Microsoft 365 and Fabric IQ. In Excel and PowerPoint, it performs at the level of GPT-6 Luna according to internal benchmarks. In Word, it approaches GPT-6.1 Sol, using about 42 percent fewer input tokens per call. In Teams, it uses up to 28 percent fewer tokens than GPT-6 Luna when following instructions and calling tools. When generating DAX queries for Power BI, the model outperforms GPT-5.6 Sol, GPT-6 Astra, and Opus 5.5 in offline tests.

MAI Code 2 and Code 2 Flash are now available in GitHub Copilot. Code 2 Flash is also in limited-access preview in Microsoft Foundry. OpenRouter, Vercel, Hermes, OpenCode, and Kilo will follow soon.