Microsoft AI chief Mustafa Suleyman says the company’s MAI-Base-2 model outperformed a comparable OpenAI model while using a quarter of the computing power to train. If that comparison holds up, it would be a striking efficiency claim from a company seeking to rely less on its partner’s models.
Microsoft AI Watch analysis
What happened
Suleyman made the comparison during a Thursday town hall with Microsoft employees, Business Insider reports. He said MAI-Base-2 was stronger than the model currently powering GPT-5.6 Sol, and used one-quarter of the compute OpenAI used to train its model. He also claimed Microsoft’s model used 12 times less computation than an unspecified DeepSeek model.
Suleyman said Microsoft’s comparisons were based on technical information it has about OpenAI’s models. The report says Microsoft declined to comment and OpenAI did not respond to a request for comment.
Why it matters
Training a capable model with less computing could reduce a major cost of building AI. It would also give Microsoft a stronger case for developing its own models rather than depending so heavily on OpenAI, even as the companies remain partners.
But the headline comparison comes from Suleyman, and the report does not provide the benchmarks, model details or compute accounting needed to assess it independently. “Better” is doing a fair bit of work until those details are available.
Our read
This is a consequential claim, not yet a settled scorecard. Microsoft’s reported result would matter both for the cost of training frontier models and for the balance of its relationship with OpenAI. The useful next step is evidence that lets others compare like with like, rather than another round of victory laps.
What to watch
- Whether Microsoft publishes the benchmark results and evaluation methods.
- Which DeepSeek model Suleyman meant in the 12-times comparison.
- Whether the claimed efficiency carries through to later models and real-world use.
Discussion spark: What evidence would convince you that Microsoft’s model is genuinely better and more compute-efficient, rather than simply winning on a comparison chosen by its own executives?
Sources and evidence
not affiliated with or endorsed by Microsoft