Microsoft's new AI model, MAI-Base-2, outperforms OpenAI's GPT-5.6 on key pre-training metrics, using only 25% of the computing power, according to CEO Mustafa Suleyman. The announcement signals a growing competitive edge for Microsoft's in-house AI development.
TL;DR
- Microsoft's MAI-Base-2 model surpasses OpenAI's GPT-5.6 in pre-training efficiency, using significantly less compute power.
- This development highlights Microsoft's increasing independence from OpenAI, despite its substantial investment in the startup.
- The model also demonstrates 12 times greater compute efficiency compared to DeepSeek's models, per Suleyman.
What happened
During a town hall meeting on Thursday, Microsoft AI CEO Mustafa Suleyman revealed that the company's new model, MAI-Base-2, outperforms OpenAI's GPT-5.6 on pre-training metrics. Suleyman stated that Microsoft achieved this using only a quarter of the compute power OpenAI used for its model. Pre-training involves teaching AI models to recognize patterns in vast amounts of data, a process that typically requires substantial computing resources.
Microsoft privately tested MAI-Base-2 against OpenAI models, including one codenamed 'Spud' that powers GPT-5.6. Suleyman claimed that Microsoft's model performed better on tests measuring the AI's ability to understand and predict patterns in information. The comparisons were based on technical data Microsoft has about OpenAI's models, suggesting special access to OpenAI's performance metrics.
Why it matters
This development underscores Microsoft's growing capability in AI model development, potentially reducing its reliance on OpenAI. Despite Microsoft's significant investment in OpenAI—estimated to be worth over $100 billion—this announcement hints at a strategic shift towards developing its own AI technologies. The partnership between the two companies has reportedly frayed in the past year, with Microsoft increasingly focusing on its own AI models.
The efficiency gains in compute power could translate to substantial cost savings for Microsoft. Additionally, the model's superior performance compared to DeepSeek's offerings highlights Microsoft's competitive edge in the global AI landscape. However, the implications for Microsoft's investment in OpenAI and their ongoing partnership remain to be seen.
Key facts
- Microsoft's MAI-Base-2 model outperforms OpenAI's GPT-5.6 on pre-training metrics.
- MAI-Base-2 uses only 25% of the compute power used by OpenAI for GPT-5.6.
- The model demonstrates 12 times greater compute efficiency compared to DeepSeek's models.
- Microsoft privately tested MAI-Base-2 against OpenAI models, including 'Spud' that powers GPT-5.6.
- Microsoft announced MAI-Base-2 on Thursday.
- Mustafa Suleyman is the CEO of Microsoft AI.
- Microsoft has invested billions in OpenAI and owns a stake likely worth over $100 billion.
- The relationship between Microsoft and OpenAI has reportedly frayed in the past year.
Context
Microsoft's investment in OpenAI has been a cornerstone of its AI strategy, with the company integrating OpenAI's models into its own products like Copilot. However, this announcement suggests a pivot towards developing in-house AI capabilities, potentially reducing dependency on external partners. This move aligns with broader industry trends where tech giants are increasingly investing in their own AI research and development.
The efficiency gains in compute power are particularly notable given the high costs associated with training large AI models. By achieving better performance with less compute, Microsoft could set a new standard for cost-effective AI development. This could also pressure other AI companies to innovate in compute efficiency, potentially leading to broader industry advancements.
