Microsoft's MAI Code 1.1 Flash trails DeepSeek on price and performance

Microsoft has released MAI Code 1.1 Flash, a new code generation model designed for GitHub Copilot. According to Microsoft's claims, the model offers three key improvements: it writes better code, is 25 percent more token-efficient (meaning it uses fewer tokens to complete the same task), and costs a quarter of its June predecessor. The model was trained using "hundreds of thousands of reinforcement-learning environments in GitHub Copilot," and developers accepted 4 percent more of its output compared to the prior version.

On published benchmarks, MAI Code 1.1 Flash shows incremental progress. It achieved 72.6% on SWE-bench Verified (compared to its predecessor's 71.6%), and 62.9% on Terminal Bench 2.1 (a jump from 51.7%). These scores edge past Anthropic's Haiku 4.5 (69.8% on SWE-bench, 49.4% on Terminal Bench) and OpenAI's GPT-5.4 mini (69.2% and 60.7% respectively). However, DeepSeek-V4-Flash-0731 significantly outperforms MAI Code 1.1 Flash on Terminal Bench 2.1, achieving 82.7%—a gap of nearly 20 percentage points. DeepSeek did not publish its SWE-bench Verified result, preventing a full comparison on that metric.

Pricing tells a similar story of trade-offs. MAI Code 1.1 Flash costs $0.20 per input token and $1.20 per output token. While this is cheaper than Anthropic's Claude Haiku 4.5 ($1.00 input / $5.00 output), it remains significantly more expensive than DeepSeek-V4-Flash at $0.14 per input token and $0.28 per output token. Even accounting for token efficiency, DeepSeek's cost advantage is likely substantial. Microsoft notably buried these benchmark results in the model card rather than highlighting them in its official announcement, instead choosing to tout only vague improvement metrics like "code survival rose 4% and return visits increased 9%" over its predecessor, omitting any direct comparisons to competitors.

The release exposes a disconnect between Microsoft's public rhetoric and its actual product strategy. The company has positioned itself as an enthusiastic supporter of open-weights AI models—freely available alternatives like DeepSeek—yet continues to invest in proprietary, weaker, and pricier alternatives. This mirrors Microsoft's recent restructuring of Copilot itself, where the company swapped out OpenAI and Anthropic models for cheaper MAI alternatives to reduce costs, accepting worse performance in exchange for better margins. By defaulting to its own models within the Microsoft ecosystem and relying on the reality that most users never actively select a different AI model, Microsoft appears to be betting that it can lock up significant market share despite offering inferior performance at higher cost.