Introduction
Microsoft has released MAI Code 1.1 Flash, a code generation model designed for GitHub Copilot. The company claims the new model writes better code, is 25 percent more token-efficient, and costs a quarter of its June predecessor. However, benchmark comparisons reveal that the model trails DeepSeek's V4 Flash on both price and performance, prompting questions about Microsoft's positioning in the competitive AI landscape.
Performance Benchmarks
In internal benchmarks, MAI Code 1.1 Flash shows modest improvements over its predecessor and some competing models. On SWE-bench Verified, it scored 72.6%, edging past MAI Code 1 Flash (71.6%), Haiku 4.5 (69.8%), and GPT-5.4 mini (69.2%). However, DeepSeek did not publish a score for this benchmark.
On Terminal Bench 2.1, the results are more telling. MAI Code 1.1 Flash scored 62.9%, a significant improvement over its predecessor's 51.7% and Haiku 4.5's 49.4%, but still well behind DeepSeek V4 Flash's 82.7%. This gap indicates that DeepSeek's model is substantially more capable in real-world coding tasks.
Pricing Comparison
Microsoft positions MAI Code 1.1 Flash as a budget-friendly option, but a closer look at pricing reveals that it is still more expensive than DeepSeek's offering. For input tokens, MAI Code 1.1 Flash costs $0.20 per million, compared to DeepSeek's $0.14. With caching, the price drops to $0.02 for MAI and $0.0028 for DeepSeek. Output tokens are where the gap widens: MAI charges $1.20 per million, while DeepSeek charges only $0.28.
These prices are significantly lower than Anthropic's Claude Haiku 4.5, which costs $1.00 for input, $0.10 with cache, and $5.00 for output. However, DeepSeek's pricing undercuts Microsoft's by a substantial margin, making it a more attractive option for cost-conscious developers.
Token Efficiency and User Adoption
Microsoft claims that MAI Code 1.1 Flash is 25 percent more token-efficient than its predecessor, meaning it requires fewer tokens to generate the same output. This efficiency could translate into lower costs for users, but the company has not provided detailed data on how this translates into real-world savings.
In terms of user adoption, Microsoft reports that developers accepted 4 percent more of the model's output compared to the previous version. Additionally, the company notes a 9 percent increase in return visits, suggesting that users are finding the model more useful. However, these metrics are relative to Microsoft's own models and do not address how the model compares to competitors like DeepSeek.
Training and Development
Microsoft trained MAI Code 1.1 Flash using "hundreds of thousands of reinforcement-learning environments in GitHub Copilot." This approach leverages real-world coding scenarios to improve the model's performance. The company has not disclosed the exact training data or methodology, but the use of reinforcement learning from user interactions is a notable strategy.
Strategic Implications
The release of MAI Code 1.1 Flash comes amid Microsoft's push to position itself as an open AI champion. However, the company's decision to invest in a proprietary model that trails open-weight alternatives like DeepSeek on both price and performance seems contradictory. Microsoft has repeatedly expressed admiration for open-weight models, yet its own strategy appears to favor in-house, closed solutions.
The likely reason is cost-cutting. Microsoft recently shuffled its Copilot offerings, replacing OpenAI and Anthropic models with its own cheaper MAI alternatives to reduce expenses. The trade-off is worse performance for better margins. MAI Code 1.1 Flash fits this pattern: it is cheaper than its predecessor but still more expensive than DeepSeek, and it performs worse.
Microsoft's customers in the GitHub Copilot ecosystem can still choose from different models depending on the app and use case. However, Microsoft will likely make its own models the default option, locking users into a less competitive solution. This strategy may alienate developers who prioritize performance and cost.
Conclusion
Microsoft's MAI Code 1.1 Flash represents an incremental improvement over its predecessor, but it fails to compete with DeepSeek's V4 Flash on either price or performance. The company's focus on proprietary models contradicts its public stance on open AI and may ultimately hinder its competitiveness in the rapidly evolving AI market. As developers increasingly seek cost-effective and high-performing solutions, Microsoft may need to reconsider its approach or risk losing ground to more agile competitors.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
