Microsoft's MAI Code 1.1 Flash Loses on Price and Performance to DeepSeek-V4-Flash
Microsoft's MAI Code 1.1 Flash beats its predecessor and mini-models from Anthropic and OpenAI on SWE-bench Verified, but DeepSeek-V4-Flash-0731 outperforms it on Terminal Bench 2.1 (82.7% vs 62.9%) while costing roughly a third as much per token.
MAI Code 1.1 Flash — Quick Specs
What happened
Microsoft released MAI Code 1.1 Flash, an updated code model powering GitHub Copilot. According to Microsoft, the model writes better code, is 25 percent more token-efficient, and costs a quarter of its June predecessor, MAI-Code-1-Flash. Microsoft says developers accepted 4 percent more of its output than the previous version, and that training used "hundreds of thousands of reinforcement-learning environments in GitHub Copilot."
The benchmarks
On SWE-bench Verified, MAI-Code-1.1-Flash scores 72.6%, ahead of MAI-Code-1-Flash (71.6%), Claude Haiku 4.5 (69.8%), and GPT-5.4 mini (69.2%). DeepSeek did not publish a comparable SWE-bench Verified score for its V4-Flash-0731 release.
On Terminal Bench 2.1, the picture reverses sharply. MAI-Code-1.1-Flash posts 62.9%, ahead of its predecessor (51.7%) and Haiku 4.5 (49.4%), and roughly in line with GPT-5.4 mini (60.7%). But DeepSeek-V4-Flash-0731 scores 82.7% on the same benchmark — nearly 20 points higher than Microsoft's model.
Pricing gap
DeepSeek-V4-Flash charges $0.14 per 1M input tokens ($0.0028 with caching) and $0.28 per 1M output tokens. MAI Code 1.1 Flash charges $0.20 per 1M input tokens ($0.02 with caching) and $1.20 per 1M output tokens — more than four times DeepSeek's output price. Claude Haiku 4.5 is the most expensive of the three, at $1.00 input and $5.00 output per 1M tokens.
Cost per token alone doesn't capture full usage efficiency, but the gap favoring DeepSeek is large enough that token-efficiency differences are unlikely to close it.
Notable omission
Microsoft's official announcement for MAI Code 1.1 Flash does not include direct benchmark comparisons against competitors. It cites internal metrics instead — "code survival rose 4% and return visits increased 9%" relative to the prior version — while the head-to-head benchmark table against Haiku 4.5, GPT-5.4 mini, and DeepSeek-V4-Flash-0731 appears only in the model card.
What this means
Microsoft has been vocal about supporting open AI development, but its own roadmap points toward proprietary, closed-weight models like MAI Code 1.1 Flash that underperform freely available alternatives such as DeepSeek-V4-Flash-0731 on some benchmarks while costing more per output token. This follows Microsoft's recent Copilot changes, where it replaced some OpenAI and Anthropic models with cheaper in-house MAI alternatives, trading capability for margin.
The pattern suggests strategy rather than accident: Microsoft controls the default model choice inside its own ecosystem, and most Copilot users never actively select a different model. Even a model that trails on independent benchmarks can capture significant usage share simply by being the default — regardless of whether cheaper, better-performing open alternatives exist elsewhere.
Related Articles
Microsoft Merges Coding and Productivity Copilot Into Single App to Counter Anthropic
Microsoft launched an updated Copilot app that merges coding, productivity tasks, and custom agent creation into three tabs — Cowork, Code, and Autopilot. The company is shifting to usage-based pricing as it tries to close the gap with Anthropic's Claude in enterprise AI adoption.
Microsoft releases Decision-1, a Qwen3.5-9B-based model for classification and routing, at $0.042 per 1M input tokens
Microsoft has released Decision-1, a decision model built on Qwen3.5-9B for classification, evaluation, and routing. Microsoft claims 83.5% accuracy across 36 benchmarks and 85 ms latency. Input tokens cost $0.042 per million, and output tokens are free.
Microsoft releases FrogNano-4B, an Apache 2.0 coding agent trained with RL on 1,500 synthetic tasks
Microsoft has released FrogNano-4B-2609, a repository-level coding agent derived from Qwen3.5-4B and published under Apache 2.0 with open weights. Microsoft says it was post-trained only with reinforcement learning on about 1,500 synthetic software-engineering tasks, with no stronger-model trajectories. It is evaluated at roughly 131K tokens of context.
Microsoft's Copilot gets access to local Windows files and OS-level actions under 'Hybrid Intelligence'
Microsoft announced an upgrade to Copilot at its Windows and Surface event that gives the assistant access to local files and the ability to take actions across Windows. The company calls the underlying approach "Hybrid Intelligence," which combines local and cloud AI models. Pricing, model details, and availability were not disclosed in the available reporting.
Comments
Loading...