Microsoft's new MAI Code 2 delivers 98% of GPT-6.1 Sol performance at 40% of the cost, alongside the high-speed MAI Code 2 Flash.

At the Microsoft Ignite conference on November 17, 2026, Microsoft unveiled a major leap in specialized generative AI with the launch of the MAI (Microsoft AI) Code series. The flagship model, MAI Code 2, is designed as a general-purpose coding model specifically optimized for everyday enterprise software engineering workflows.
The release is part of a massive seven-model MAI launch, signaling Microsoft's aggressive push to dominate the developer tooling market. While the industry has been chasing raw reasoning power, Microsoft is pivoting toward a strategy of extreme efficiency, offering high-tier intelligence at a fraction of the current market rates for frontier models.
MAI Code 2 is a proprietary, closed-source model that was developed entirely in-house, with both pre-training and post-training conducted on Microsoft's specialized infrastructure. Unlike many contemporary models that rely on hybrid data sources, MAI Code 2 benefits from a highly curated dataset focused on high-quality code repositories and technical documentation.
A significant architectural highlight is the MAI Code 2 Flash variant. This model was created using advanced distillation techniques, where the capabilities of the larger MAI Code 2 'teacher' model were transferred into a smaller, more agile 'student' model. This allows Flash to maintain high software engineering proficiency while providing the low latency required for real-time autocomplete and IDE integrations.
The performance metrics for MAI Code 2 position it as a formidable competitor to the industry's most expensive models. When averaged across four rigorous benchmarks—DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas, and FrontierCode v1.1—the model achieved a 63.2% pass rate at maximum reasoning effort. This represents 98% of the performance seen in GPT-6.1 Sol, but at only 40% of the cost.
Perhaps most impressive is the comparison to Claude Opus 5.5. MAI Code 2 reaches 96% of the Claude Opus 5.5 average (65.3%) while operating at a staggering 5% of the cost. In blind frontend evaluations, human reviewers actually preferred MAI Code 2's output over GPT-6 Sol in 64% of cases, highlighting its practical utility in real-world UI/UX coding tasks.
The MAI Code 2 Flash model also shows remarkable strength, averaging a 61.0% pass rate. Notably, it outperforms Claude Haiku 5.5 by 12% while costing only one-fifth as much.
Microsoft is leveraging aggressive promotional pricing to encourage enterprise migration. For developers using the model via API, the current rates are significantly lower than standard market pricing. These promotional rates are scheduled to remain in effect until May 17, 2027, after which the list price is expected to approximately double.
The pricing structure is particularly advantageous for workflows involving repetitive context, such as large codebase analysis, thanks to highly discounted cached input rates. This makes MAI Code 2 one of the most cost-effective options for RAG-based coding assistants and automated PR review agents.
MAI Code 2 is engineered for the daily grind of professional software development. It excels at code completion, refactoring, unit test generation, and explaining complex legacy codebases. Its high performance in Terminal-Bench 2.1 makes it an ideal candidate for CLI-based automation and DevOps scripting.
While it is not currently optimized for fundamental scientific research or extremely long-horizon agentic tasks (tasks requiring hundreds of sequential steps), it is perfect for 'human-in-the-loop' development. Developers can rely on it for high-accuracy code suggestions within GitHub Copilot or for building custom internal tools via Microsoft Foundry.
Developers can access MAI Code 2 immediately through two primary channels. For individual developers and teams already using GitHub, the model is integrated directly into GitHub Copilot, providing a seamless experience within VS Code and other supported IDEs.
For enterprise architects and AI engineers looking to build custom applications, the model is available via Microsoft Foundry. This allows for deeper integration into existing cloud infrastructures and more granular control over model deployment and fine-tuning workflows.
API Pricing — Input: 0.50 USD per million tokens (promotional, until May 17 2027) / Output: 2.50 USD per million tokens (promotional) / Context: Includes 0.025 USD per 1M tokens for cached input (promotional until May 17, 2027)