Microsoft has introduced MAI Code 2, a new AI model designed for programming. The company says its code-generation quality is approaching that of leading solutions from OpenAI and Anthropic, while costing significantly less.

Microsoft is betting not on absolute benchmark leadership, but on the balance between performance and cost.

Benchmark results and cost comparison

According to Microsoft’s internal tests across four coding benchmarks (DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas, and FrontierCode v1.1), MAI Code 2 scored an average of 63.2%, narrowly trailing GPT-6.1 Sol (64.1%) and Claude Opus 5.5 (65.3%). In percentage terms, the new model delivers about 98% of OpenAI’s solution’s performance and 96% of Anthropic’s model’s capabilities.

MAI Code 2

Microsoft’s main argument for the model is its cost efficiency:
Compared with GPT-6.1 Sol. Completing a typical task costs an average of $0,309, compared with $0,761 for OpenAI’s model—a reduction of about 60%.

Compared with Claude Opus 5.5. The price difference is as much as twentyfold: $0,309 versus $6,306 per task (about 5% of the competitor’s cost).

MAI Code 2

However, this aggressive pricing comes with an important caveat: the quoted rates include a temporary 50% discount, which is available until May 17, 2027. After the promotional period ends, standard rates will double (to $1 per million input tokens and $5 per million output tokens), but even then MAI Code 2 will remain about ten times cheaper than Claude Opus 5.5.

What the model is designed for and where it’s available

Microsoft emphasizes that MAI Code 2 was designed for day-to-day enterprise development: deep debugging, editing multiple files at once, analyzing large repositories, and building user interfaces. In blind testing of frontend tasks, developers preferred the dashboards and visualizations generated by MAI Code 2 over GPT-6 Sol’s results in 64-68% of cases.

MAI Code 2

Alongside the standard version, Microsoft released a lightweight model, MAI Code 2 Flash, for quick, targeted edits and code explanations. It competes with GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their respective costs.

The company also candidly acknowledges the new models’ limitations: they are not yet optimized for fundamental scientific research or complex autonomous AI agents with long planning horizons, areas where more powerful solutions such as Claude Opus 5.5 remain ahead. On the SWE-Atlas benchmark, Anthropic’s model leads Microsoft’s new model by 14.4 percentage points.

Both models are already integrated into GitHub Copilot and available to paid subscribers, while the Flash version is also available for preview testing on the Microsoft Foundry platform.