Suleyman Told Staff MAI-Base-2 Beats the Model Behind GPT-5.6 Sol on a Quarter of the Compute. Microsoft's Own Paper Only Compares Open Models.

Microsoft made two claims about the same model on the same day. One comes with a 13-page PDF, a parameter table and per-benchmark scores. The other comes from an employee town hall, and it is the one getting the headlines.

On Thursday, October 15, 2026, Microsoft AI published Frontier Pre-training at Microsoft AI, saying it had finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active per token. The post says MAI-Base-2 "outperforms every open-weight model we tested on our held-out evaluations."

The same day, according to Business Insider, Microsoft AI CEO Mustafa Suleyman told employees at a town hall that Microsoft now has a stronger pre-trained base model than the one powering OpenAI's GPT-5.6 Sol, and that it used one-quarter of the compute OpenAI used to train its model. Reports relaying Business Insider, including Windows Mode and DigitalToday, say he described the OpenAI comparisons as private tests based on technical information Microsoft holds about OpenAI's models.

None of that comparison is in the paper. The word OpenAI does not appear in it. That gap is the story, and it is also why the paper is worth reading closely.

What Microsoft Actually Published

Microsoft trained MAI-Base-2 on 16,000 Nvidia GB200 GPUs over 38 days in August and September 2026. It reports average model FLOPs utilization of 21 percent and goodput, the share of time training actually progressed, of 89 percent. That is roughly 14.6 million GPU hours by our arithmetic. Here is how the new model compares with the MAI-Base-1 model it replaces, using Microsoft's own table.

A few details stand out. Vision data is 18 percent of the 50 trillion token mixture, drawn from a 4.5 trillion token unique vision corpus. The unique-token corpus grew about 36 percent over MAI-Base-1, and web data has a March 2026 knowledge cutoff. Fixing two MoE inefficiencies raised the MFU of Microsoft's MoE layers from 24 to 35 percent, which is what let it put MoE in every layer.

The model is sparse even by current standards. About 2.9 percent of its weights are active per token, down from 3.5 percent for MAI-Base-1.

The Benchmarks Microsoft Chose

Microsoft scores base models on bits per byte (BPB) over held-out text, where lower is better. The code set is Microsoft's own codebase; the STEM, math and Q&A sets were commissioned from vendors to limit contamination. These are the values printed in Figure 1 of the PDF, with parameter counts as labeled there.

*Post-trained checkpoints. Microsoft notes that only post-trained versions of these models are available and that post-training may affect the comparison.

MAI-Base-2 wins all four columns. Against DeepSeek-V4-Pro-Base, the only true base model in the comparison besides Microsoft's own, the margin is 4.3 to 7.9 percent depending on the set. Against Kimi K3 it is narrower: 2.5 to 2.9 percent on code, Q&A and STEM, and 0.8 percent on math.

From those fits Microsoft derives its efficiency headline. It estimates its recipe is 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base, meaning a model on its scaling ladder would match DeepSeek's scores with that much less training compute. It also claims a 6.3 to 12.7 times gain over MAI-Base-1 since that model's June release. The 12 times figure Suleyman reportedly used for DeepSeek sits at the top of that published range.

What a Quarter of the Compute Would Mean

Microsoft did not publish a total FLOPs figure, so here is our own back-of-envelope estimate. The standard approximation of six times active parameters times tokens gives 6 × 80 billion × 50 trillion, or about 2.4 × 10²⁵ FLOPs. A cross-check from the run statistics (16,000 GPUs at 2.5 petaFLOPS BF16 peak, 21 percent MFU, 38 days) lands at about 2.8 × 10²⁵. The two agree closely.

If Suleyman's one-quarter figure is measured the same way, it would put the pre-training run behind GPT-5.6 Sol somewhere near 10²⁶ FLOPs. That is our inference, not anything Microsoft or OpenAI has said. We do not know whether Suleyman meant FLOPs, GPU hours or dollars, and 75 percent less compute is not the same as 75 percent less spend.

Why Microsoft Is Saying This Now

Microsoft still holds a large stake in OpenAI. In October 2025 it valued that stake at approximately $135 billion, and the same agreement said Microsoft "can now independently pursue AGI alone or in partnership with third parties." The April 2026 rewrite, as DatacenterDynamics reported, left Microsoft with an IP license to OpenAI models through 2032 that is now non-exclusive.

Since then the in-house push has been steady. In July, Bloomberg reported Microsoft was routing Excel and Outlook prompts to its own MAI models. On October 12 we covered its plan to make Maia its primary inference system. A frontier-scale base model is the missing piece between Microsoft's own chips and Microsoft's own products.

The paper also makes a commercial pitch. Microsoft says its pre-training "does not involve distillation from other models," that it processed every training token in-house, and that this provenance will let it "license our model weights to trusted partners through Foundry." It says it tunes models first for its own products, then for Foundry customers. That is a sales argument aimed at regulated enterprises, and it is a sharper one than any benchmark.

Caveats

MAI-Base-2 is not a product. Microsoft says it is still in post-training and gives no release date, API or price. BPB measures how well a base model predicts held-out text; it says nothing directly about chat quality, tool use or agent reliability, which is where GPT-5.6 Sol competes. Every number in the paper is self-reported, and three of the five rivals are post-trained models scored on a base-model metric. The OpenAI claim rests on reports of an internal meeting. OpenAI has not publicly confirmed or disputed it, and Microsoft has not published the comparison.

Our Take

The paper is the real news, and it is a good paper. A US lab with a closed flagship published its GPU count, run length, MFU, goodput, token count, architecture and per-benchmark scores. That is a level of run-level detail we rarely see outside open-weight labs, and it makes Microsoft's open-model claims checkable the moment anyone gets the weights.

The GPT-5.6 line is a different kind of claim. It was made to employees, it relies on data about a partner's models that no outsider can see, and it compares base models when customers only ever touch post-trained ones. We think Suleyman may well be right on pre-training loss. We also think that is the easier half. In our view, OpenAI's lead has come as much from post-training and product as from base models, and Microsoft has not shown its post-training yet.

For builders, the practical read is simple. Do not plan around MAI-Base-2 until a post-trained model ships with public benchmarks and a Foundry price. Do watch the license terms. Weights that a Fortune 500 legal team can host, with a clean data-provenance story, would be a more disruptive product than a few points of BPB.

Three signposts for the next 60 days: whether Microsoft ships a post-trained MAI-Base-2 model with scores on public agent and coding benchmarks, not just BPB; whether Foundry names its first weight licensees and their terms; and whether OpenAI answers the compute claim with numbers of its own.

Sources: Microsoft AI (Frontier Pre-training at Microsoft AI), Microsoft AI PDF, Business Insider, Windows Mode, DigitalToday, 24/7 Wall St. via Yahoo Finance, Microsoft (October 2025 partnership post), DatacenterDynamics.