Microsoft releases its AI-based code generator. But it gets beaten by DeepSeek!

Microsoft published MAI Code 1.1 Flash, a state-of-the-art coding model for GitHub Copilot. According to Microsoft, it’s 25% more token efficient than its previous version and one quarter of the price. These are good news, however, when it comes to the competition, Microsoft is still behind.

1What the Benchmarks Say

MAI Code 1.1 Flash achieves 72.6% on SWE-bench Verified, a benchmark for solving real world problems based on GitHub repositories. While this result is better than some other models (69.2–69.8%), it’s still far behind DeepSeek-V4-Flash which achieves 82.7% on the more challenging Terminal Bench 2.1. MAI Code 1.1 Flash gets 62.9% on this benchmark.

Moreover, the price is a huge factor here. Microsoft charges $0.20 per thousand input tokens while DeepSeek only asks for $0.14. Even with caching, the price ratio stays the same: $0.02 vs. $0.0028 (7x). For output tokens, it’s $1.20 vs. $0.28.

2What This Means for the Market

Again, closed source models are beaten by open source ones both in terms of performance and price. The competition in the AI assisted code generation space is heating up. Companies have to invest in their own brand and ecosystem, yet they are outperformed by open source models in benchmarks. It will get harder for Microsoft and other large corporations.