Suleyman says Microsoft's base model beat OpenAI's on a quarter compute
Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls.
Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model FLOPs utilization and 89% goodput — the fraction of wall-clock time training actually progressed. Microsoft says the model beat every open-weight model it tested on its own held-out evaluations, and that the efficiency saving was spent on training a larger model rather than banked.
Then the framing. At a Thursday town hall, Microsoft AI CEO Mustafa Suleyman told employees the company "now has a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol," and that it used "one-quarter of the compute that OpenAI used to train their model." He put the DeepSeek comparison at 12 times less computation. Microsoft declined to comment and OpenAI did not respond to a request for comment, so these are one executive's claims, made privately and then reported. They also rest on technical information Microsoft holds about OpenAI's models that nobody outside Redmond can see — which is precisely why the comparison cannot be independently checked. For context, Business Insider reports Microsoft's stake in OpenAI Group PBC is likely worth more than $100 billion.
Two things keep this from being a product launch. MAI-Base-2 is a base model: pre-training is done, post-training is in progress, and there is no released checkpoint or price. And Microsoft is explicit that the weights are headed for Foundry, licensed to enterprises rather than posted publicly, trained on what it calls organic data with no distillation from other models. A licence-gated model with a provenance claim is a different pitch to enterprises than an open-weight drop — and it keeps the model inside Microsoft's own products first, with MAI-Base-3 prep already under way. Read the efficiency claim as a strategy statement as much as a benchmark result: if the recipe really is several times cheaper per unit of quality, the binding constraint moves off raw compute and onto data pipelines and infrastructure tuning, which is the part Microsoft says it now owns end to end.
The White House's Super Intelligence task force formally adopted its charter on Thursday, starting a 120-day clock on a report about how the government should handle frontier AI. Director of National Intelligence Jay Clayton chairs the group with National Cyber Director Sean Cairncross. Pentagon CTO Emil Michael, a vice chair, gets the job of building protocols for labs to report breaches, hacks and other cyber incidents, and of running the interface with frontier labs; FTC chair Andrew Ferguson leads a law enforcement and consumer protection committee with the Justice Department; OPM director Scott Kupor takes critical infrastructure plus a Labor Department study of AI's effect on the workforce.
The headline deliverable is a risk assessment from the President's Council of Advisers on Science and Technology covering cybersecurity and critical infrastructure, biosecurity and dual-use applications, and model control and accountability. A working group will implement the White House Accord on Super Intelligence that frontier lab CEOs signed this month. The gap is enforcement: nothing announced on Thursday says how compliance gets checked, and the administration's stated preference remains voluntary audits over new rules. We covered the appointment when Trump named Clayton to run it — Jay Clayton will run Trump's Super Intelligence Force — and the deadline was always the easy half of the job.
What to watch: whether MAI-Base-2's efficiency curve survives contact with anyone outside Microsoft — the first post-trained models built on it, and whether an independent run reproduces the FLOPs-efficiency gap.
Does an efficiency number you can't check change how you read a model launch — or do you wait for the checkpoint you can actually download? Tell us in the comments.
