Microsoft has begun a direct assault on Anthropic's core territory.
At its annual Build developer conference held on Tuesday, Eastern Time, the company unveiled a suite of seven new AI models spanning reasoning, coding, vision, and multimodal capabilities. The most notable announcements were Microsoft's first flagship reasoning model, MAI Thinking-1, and a cost-effective, high-efficiency coding model for GitHub scenarios, MAI-Code-1-Flash.
In an interview, Microsoft AI head Mustafa Suleyman stated the company is pursuing a different path from Google, Meta, and OpenAI, focusing more on an "Anthropic-style" market of enterprises, developers, and programming. He remarked, "We are more focused on the Anthropic direction—enterprises, developers, and coding."
This series of model launches also signals a new phase in Microsoft's strategy for AI "autonomy": while maintaining its deep partnership with OpenAI, Microsoft is accelerating the development of its own cutting-edge model portfolio.
A Broad Range of New Models
According to Microsoft's announcement, the seven new models are part of the expanding Microsoft AI (MAI) family, covering various capability levels and use cases.
The key products include the MAI Thinking series of reasoning models, ultra-efficient code models, vision and multimodal models, lightweight models for agent systems, and models optimized for enterprise and developer use.
Microsoft describes this as part of building a "hill-climbing machine"—a model development system that continuously iterates and self-improves.
The goal of the new model suite is not merely to chase parameter scale, but to build a capability stack supporting the next generation of AI Agent systems, encompassing thinking, reasoning, execution, and coding.
MAI Thinking Targets Anthropic's Claude Sonnet 4.6
The most significant launch is Microsoft's first reasoning model family, MAI Thinking.
Reasoning models have become a focal point of AI competition. Compared to traditional chat models, they emphasize multi-step deduction, complex task decomposition, long-chain planning, mathematical and code reasoning, and agent task execution.
Microsoft states that MAI Thinking can break down complex problems into smaller, more manageable steps and is specifically optimized for coding, developer workflows, and agent tasks.
According to Microsoft's data, models in the MAI-Thinking series achieve coding capabilities close to the level of Anthropic's Claude Sonnet 4.6, released in February.
Microsoft's announcement notes that MAI-Thinking-1 is the flagship reasoning model within Microsoft AI. As a medium-sized model, it is top-tier in its class, performing on par with leading models in key software engineering benchmarks and matching Sonnet 4.6 in human preference evaluations.
Suleyman acknowledged to media that Anthropic still holds a lead of several months, but emphasized that Microsoft is rapidly closing the gap. He stated that Microsoft is now at the absolute forefront and has closed a significant gap within six months.
This stance reflects a shift in Microsoft's current AI R&D strategy—no longer content to merely utilize OpenAI models, but aiming to compete directly in the front lines of model development.
Ultra-Efficient Coding Model: Targeting GitHub and Enterprise Development
In addition to reasoning models, Microsoft released an "ultra efficient" coding model, MAI-Code-1-Flash. Reportedly fine-tuned specifically for the GitHub platform, Microsoft describes it as a reasoning-efficient, agent-style programming model deeply integrated with GitHub Copilot, VS Code, and the Microsoft technology stack. With 5 billion parameters, its performance is comparable to Haiku's but at a lower cost.
Coding models have become one of the clearest paths for AI commercialization. Current competitors include Anthropic's Claude Code/Cowork, OpenAI's Codex family, Google's Gemini Code capabilities, and AI-native development platforms like Cursor and Replit.
Microsoft clearly aims to leverage the GitHub ecosystem to reinforce its advantages.
Suleyman believes the combination of reasoning and coding models will be key to the next stage of AI agent development. He said the "thinking + coding" capability combination will help Microsoft build true agent systems—intelligent entities capable of autonomously completing tasks.
Why Microsoft is Focusing on Anthropic
Notably, Microsoft executives publicly positioned Anthropic as their primary benchmark.
Suleyman explicitly told the Financial Times that Microsoft is "not that interested" in the more consumer-oriented routes of Google, Meta, and OpenAI.
Microsoft is more focused on the enterprise developer market that Anthropic targets. The reason is straightforward.
Anthropic's rapid rise over the past year has seen its strengths increasingly concentrate on Microsoft's core commercial territory: enterprise software, AI programming, developer tools, and white-collar office automation.
Particularly, Anthropic's launch of Cowork, an AI coding and office tool for enterprise users, has raised concerns in the enterprise software sector. Related products have even triggered broad declines in software stocks, putting pressure on Microsoft's share price, which has fallen over 6% year-to-date, underperforming the S&P 500's gain of over 10%.
For Microsoft, this is no longer just a competition in model capabilities but a defensive battle for its moats around Office, GitHub, Copilot, and enterprise software.
Accelerating Independence from OpenAI
At a deeper level, these model launches are part of Microsoft's advancing AI autonomy strategy.
In recent years, many of Microsoft's AI products have heavily relied on OpenAI models. From Copilot to Azure AI to enterprise services, OpenAI has been the core source of Microsoft's AI capabilities. However, this dynamic is changing.
Reports indicate that after readjusting its partnership with OpenAI last year, Microsoft began actively promoting a "truly self-sufficient" path.
Microsoft still holds about a 27% stake in OpenAI and retains long-term access to advanced models, but the company has clearly started building a multi-model strategy to reduce single-source dependency.
Over the past year, Microsoft has taken a series of steps: forming a super-intelligent research team, strengthening internal model R&D, investing in cloud cooperation with Anthropic, establishing the MAI model system, and advancing agent frameworks and infrastructure.
Suleyman noted that having in-house models will also bring direct financial benefits.
Currently, when providing certain AI services to customers, Microsoft must pay partners "significant profit shares." Mature internal models could significantly reduce this cost pressure. Suleyman said, "This will be directly reflected on the P&L."
Comments