
The Zhitong Finance App learned that Microsoft (MSFT.US) is speeding up the layout of self-developed AI models. Mustafa Suleyman (Mustafa Suleyman), the head of Microsoft's AI, said that Microsoft is making significant progress in building its own cutting-edge general AI model, and also launched six special AI models at the Ignite conference, and claims that these models can achieve comparable performance at a cost far lower than that of leading OpenAI and Anthropic products. According to Microsoft, some tasks that originally took weeks to complete can now be completed in a few minutes.
Core highlights: From relying on external models to self-developed competition
Microsoft's most advanced models have long relied on OpenAI and other AI companies. This announcement shows that Microsoft can develop its own competitive technology, and it also shows that it plans to actively compete in price. Microsoft unveiled six new AI models at the Ignite conference in San Francisco on Tuesday, covering coding, cybersecurity, image generation, and voice.
Take a look at it specifically:
MAI Code 2: Microsoft says its coding model reached 98% of GPT-6.1 Sol's performance, and the price was 40% of the latter; it reached 96% of Anthropic Claude Opus 5.5's performance, and the cost was only 5% of the latter. Designed to handle complex, time-consuming programming tasks, this flagship coding model works across multiple files and understands entire software projects.
MAI Code 2 Flash: Smaller, faster model. Microsoft claims that its performance matches the GPT-6 Luna, and the price is 65% of the latter; if it matches Claude Haiku 5.5, the cost is 20% of the latter.
MAI Cyber 2 Flash: Cybersecurity model. Microsoft claims it beat GPT-6 Astra on a certain security benchmark, costing less than 2% of the latter per task. Microsoft is also planning to provide a network-specific model to a limited customer base to help deal with the surge in software vulnerabilities increasingly revealed by AI tools.
Other models: According to Microsoft, MAI Image 2.7 is number one in image editing; new transcription and real-time speech models are designed to challenge leading rivals.
Strategy: Exchange a dedicated model for a cost advantage
Cost pressure from enterprise customers is an important background to this announcement. Suleiman said he was constantly hearing complaints from companies that “token spending is soaring.” He said, “Our promise is to provide the best price/performance ratio on the market so you can actually expand your agency; we also know that you want a model that really belongs to you.”
Microsoft's strategy is to build dedicated models for specific tasks to achieve comparable performance at significantly lower costs. Suleiman said, “It basically means you're cramming more tokens, knowledge, and expertise into the same number of parameters.” He also said, “This isn't just a model; it's not just a flash. It's a system we built, a machine that climbs continuously, and collaboratively optimizes performance and price.”
But this strategy has a trade-off: models optimized for everyday coding tasks aren't as strong in advanced mathematics or scientific research. Additionally, Microsoft's performance is more dependent on selected benchmarks, and some price advantages also reflect promotional discounts.
Enterprise platform and cost control
In addition to the model itself, Microsoft also emphasizes enterprise-grade platforms and cost control tools on Ignite. Executives focused on the Microsoft IQ platform, which aims to help companies organize and share employee knowledge, expertise, and tools, so that AI systems can learn how the company operates, and continue to accumulate on this basis.
Microsoft CEO Satya Nadella (Satya Nadella) emphasized that the new AI model covers scenarios such as coding, cybersecurity, and decision-making. He also highlighted Copilot Fusion, which he said can orchestrate multiple models to deliver cutting-edge results at a significantly lower cost. Microsoft also showcased a model selection feature that can select the right model for the task and manage the cost per request while maximizing output quality.
Industry background: Frontier AI competition heats up at the same time
While Microsoft is targeting industry leaders, cutting-edge AI itself is also rapidly evolving. OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday. Among them, the price of the new Astra is 20% lower than the previous version. At the same time, Microsoft is also promoting a cutting-edge general language model to compete on a wider range of tasks.
Suleiman said that Microsoft has achieved the most advanced pre-training results and is making strong progress in reinforcement learning, but declined to give a release schedule for the model. “My team is building real superintelligence,” he said.
Potential impact: Microsoft's relationship with OpenAI or reshaping
The new model could reshape Microsoft's relationship with OpenAI and other external AI providers. MAI Code 2 will be available in GitHub Copilot and may shift some coding workloads from external models to Microsoft's own models. When asked how much Copilot traffic Microsoft expects to shift to its own model, Suleiman declined to give a target, but predicted that the adoption rate would be significant.
In terms of enterprise business, Microsoft also provides licensing model weights to enterprise customers, allowing customers to customize their models with proprietary data, but requires deployment within Azure. Suleiman said this approach “feels like open source” while providing greater security and control.
Safety and alignment are prioritized
Suleiman said recent cybersecurity incidents — taking OpenAI's AI agent hacking the Hugging Face platform as an example — prompted him to focus not only on performance, but also on key safeguards. Preventing AI from taking harmful actions and ensuring that it behaves as intended, that is, “aligning” research, is his top priority.
At the opening of the conference, Nadella also acknowledged the cross-regional service failure that occurred last week, saying that the company takes customer trust seriously, has work to be done, and promises continuous learning and improvement.
Next, the market is concerned about two points: one is how fast Microsoft can advance in the field of dedicated AI into a competitive general model; the other is whether its cost advantage is enough to push customers to switch from mature competitors to Microsoft.
Microsoft closed down 2.27% on Tuesday, with a cumulative increase of 5.3% so far this year, underperforming the S&P 500. As of press time, the stock was up 0.9% after the market.