Microsoft’s MAI-Base-2 Is Here: Suleyman Says It Beats OpenAI’s GPT 5.6 Model With a Quarter of the Compute

Microsoft AI has finished pre-training MAI-Base-2, its largest in-house foundation model so far. In a technical post published on October 15, 2026, the team says the model has 2.8 trillion total parameters (80 billion active) and beats every open-weight model it tested. Later that day, Microsoft AI CEO Mustafa Suleyman went further, telling employees the model is stronger than the base model behind OpenAI’s GPT 5.6 Sol while using a quarter of the compute.

MAI-Base-2 follows MAI-Base-1, the base model behind MAI-Thinking-1, which Microsoft announced in June. A base model is the raw result of pre-training: the stage where an AI learns patterns from huge amounts of text, images, and code. It is not a finished assistant yet. Post-training, which is happening now, is what turns it into something you can chat with.

Here are the key numbers from Microsoft’s report:

• Size: 2.8 trillion total parameters, with 80 billion active for each token.
• Hardware: trained on 16,000 NVIDIA GB200 GPUs over 38 days in August and September 2026.
• Data: a 50 trillion token training mix, with a March 2026 knowledge cutoff. Vision data made up 18% of it.
• Native vision: unlike MAI-Base-1, MAI-Base-2 was trained with images from the start, including charts, documents, and computer use tasks.
• No distillation: Microsoft says the model learned from “organic data” and was not trained on output from other companies’ models.

How Microsoft Measured It Benchmarks

Microsoft compared base models using “bits per byte” on held-out tests in code, STEM, math, and Q&A. Lower is better. MAI-Base-2 scored lower than DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B, and MiMo-V2.6-Pro-RL. It also says its training recipe has improved 6.3 to 12.7 times in compute efficiency since MAI-Base-1 in June.

According to Business Insider, Suleyman told employees at a town hall on Thursday: “We now have a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol. And in fact, we used one-quarter of the compute that OpenAI used to train their model.”

He said Microsoft tested the model “privately and confidentially” against OpenAI models, including one nicknamed “Spud” that powers GPT-5.6, using technical information Microsoft has about OpenAI’s models. He also said the model used “12 times less computation” than DeepSeek, which lines up with the top of the 4.2 to 12.3 times range in the public report.

Why it is a big deal: Microsoft has invested billions in OpenAI and sells OpenAI models to Azure customers. A Microsoft-built model that matches or beats OpenAI’s would give it far more control over the AI inside Copilot, Windows, and Azure.

Foundry Gets a Microsoft-Made Frontier Model Azure

Microsoft says it will license MAI-Base-2’s weights to “trusted partners through Foundry.” That is notable. Today, the biggest models in Microsoft Foundry come mostly from partners like OpenAI. A first-party model with known training data could appeal to companies in regulated industries that want to know exactly where their AI came from.

Lower Costs Could Mean Cheaper AI Features Copilot

Microsoft says owning the full stack lets it tune models and infrastructure together “for the best performance at the lowest cost, first in our own products and then for Foundry customers.” In plain terms, Microsoft’s own apps, likely including Copilot, should get these models first. Cheaper training and serving could also help Microsoft offer more AI features without raising prices, though it has made no promises.

What Developers Should Do Now Developers

• Nothing to call yet: there is no public API, model card, or pricing. Watch the Foundry model catalog for MAI releases.
• Plan for choice: if you build on Azure OpenAI today, keep your code model-agnostic so you can test MAI models when they arrive.
• Watch the license terms: “trusted partners” suggests licensed weights may be limited to approved enterprise customers at first.

Microsoft says more results will come once post-training is done, and work on MAI-Base-3 has already started. The model also builds on Microsoft’s push for its own AI hardware and datacenters, which we covered in our look at Azure Maia 300 and Cobalt 200.

Sources: Microsoft AI: Frontier Pre-training at Microsoft AI and Business Insider.