Microsoft AI Claims to Outperform OpenAI with Less Computing Power

Key Takeaways

1Mustafa Suleyman claims that MAI-Base-2 surpasses the model powering GPT 5.6 Sol

2The pre-training reportedly required a quarter of OpenAI's computation and twelve times less than DeepSeek

3Microsoft remains a major partner and investor in OpenAI, despite a strained relationship

Why it matters : Microsoft is claiming technical advancements while rebalancing its dependence on OpenAI, which could influence the dynamics of the industry.

Mustafa Suleyman claims to employees that the MAI-Base-2 model surpasses the AI powering GPT 5.6 Sol, with pre-training conducted using a quarter of the computing power of OpenAI. He also asserts superior efficiency compared to DeepSeek. These statements come as Microsoft remains both a partner and a major investor in OpenAI.

Partners and Investors, Microsoft and OpenAI Tense Up

Microsoft has invested billions of dollars in OpenAI and holds a stake estimated at over $100 billion. The publisher continues to offer OpenAI's models to its Azure customers and integrate them into its products, including Copilot. At the same time, the relationship has deteriorated over the past year, with Microsoft working to diversify by developing its own models. On Thursday, Microsoft declined to comment on these matters, and OpenAI did not respond to a request for comment. A public critique of OpenAI could, according to the envisioned scenario, affect the value of Microsoft's investment.

Suleyman Claims a More Powerful Model than GPT 5.6 Sol

On Thursday, Mustafa Suleyman stated to employees that Microsoft has a pre-trained base model that is more powerful than the one powering GPT 5.6 Sol. He indicated that he had privately compared this model to OpenAI systems, including one nicknamed Spud associated with GPT-5.6, and assured better performance in understanding and pattern prediction tests. He specified that these comparisons were based on technical information held by Microsoft and mentioned evaluations conducted confidentially, suggesting privileged access to performance data.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Pre-training Presented as Much More Computationally Efficient

Mustafa Suleyman claims that the model was pre-trained with a quarter of the computing power used by OpenAI. He also asserts a computational efficiency twelve times greater than that of DeepSeek, without specifying which model from that lab was targeted. The pre-training of MAI-Base-2 is presented as significantly more efficient than a DeepSeek-V4-Pro-Base offering. Pre-training, the first phase of learning from large volumes of texts, images, and other data, requires substantial costly resources; thus, better efficiency could translate into savings for Microsoft.

MAI-Base-2 Officialized and Organizational Restructuring in March

MAI-Base-2 was announced earlier in the day. In March, Microsoft merged its consumer and commercial Copilot teams and reassigned Mustafa Suleyman to build AI models within the so-called superintelligence unit.