Microsoft has introduced MAI Code 2 and MAI Code 2 Flash, two new AI models now available in GitHub Copilot. Designed for enterprise coding workflows, they aim to deliver strong performance at significantly lower costs than leading third-party models. The rollout marks a strategic shift toward optimised, task-specific AI assistance in software development.
Performance tailored to real-world tasks
MAI Code 2 targets complex development work such as deep debugging, multi-file implementation, refactoring, and repository-wide reasoning. According to Microsoft, it achieves 98% of GPT-6.1 Sol’s performance on common enterprise coding benchmarks while costing only 40% as much. Against Claude Opus 5.5, it delivers 96% of the performance at just 5% of the cost. These comparisons are based on averaged results across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1.
The model also shows strength in frontend development. In blind evaluations, reviewers preferred its UI outputs over GPT-6 Sol 64% of the time. This suggests tangible gains in visual coding quality—useful for teams building dashboards or web interfaces.
A lightweight option for routine work
MAI Code 2 Flash is positioned as a cost-efficient alternative for high-volume, routine tasks: code explanations, targeted edits, and quick fixes. It matches GPT-6 Luna’s performance at 65% of the cost and equals Claude Haiku 5.5 at 20% of the price across three benchmarks for everyday coding.
Its broad availability reflects its role as a scalable helper: accessible to Copilot Free, Student, and all paid tiers. In contrast, MAI Code 2 is limited to Pro, Pro+, Max, Business, and Enterprise subscribers. This tiered access lets organisations balance cost and capability based on developer needs.
Integration across Microsoft’s stack
Beyond Copilot, these models are becoming foundational across Microsoft products. A further optimised version, MAI Work Flash, supports M365 and Fabric IQ. In internal tests, it matches GPT-6 Luna in Excel and PowerPoint for accuracy and token efficiency, uses 42% fewer input tokens than GPT-6.1 Sol in Word, and outperforms GPT-6 Luna in Teams for instruction following and decision tracking.
In FabricIQ, the MAI model surpasses GPT-5.6-sol, GPT-6-astra, and Opus-5.5 in generating DAX queries for Power BI. This integration signals Microsoft’s push to unify AI assistance across productivity and data tools—reducing reliance on external providers.
What it means
For UK, UAE, US, and Kazakh enterprises, the launch offers a chance to reduce AI spending without sacrificing output quality. Since MAI models are developed in-house, they also reduce dependency on third-party AI vendors. However, Microsoft notes these models are not yet tuned for long-horizon agentic work or scientific research—limiting their scope.
Organisations using Copilot at scale can now assign Flash for routine tasks and reserve Code 2 for complex engineering, optimising cost. Businesses with Copilot Enterprise or Business plans can manage model access via policy settings, enabling governance.
Developers should test both models in their workflows, particularly for debugging and UI generation. Early data suggests Code 2 may outperform rivals in real-world coding tasks, but independent validation is pending.
Microsoft’s move strengthens its AI stack while offering measurable efficiency gains. For now, the models are only confirmed in Copilot and select Microsoft products, with future availability expected on Open Router, Vercel, and others.