Microsoft AI chief says his model beat OpenAI's using a quarter the compute
Mustafa Suleyman told his own staff that Microsoft's in-house model outperforms the OpenAI system powering GPT-5.6. He says it did that with roughly 75% less compute, and the timing is the real story given Microsoft is still OpenAI's biggest financial backer.
Mustafa Suleyman, the CEO of Microsoft AI, told employees that the company's new MAI-Base-2 model beats the OpenAI model that powers GPT-5.6, according to a report from TechCrunch. The claim: MAI-Base-2 used about a quarter of the compute that OpenAI spent training its comparable system, internally nicknamed Spud. Suleyman went further and said Microsoft privately benchmarked MAI-Base-2 directly against that OpenAI model on tests measuring how well an AI understands and predicts patterns in information, and that Microsoft's model came out ahead.
He didn't stop at OpenAI. Suleyman also positioned MAI-Base-2 against DeepSeek, the Chinese lab that rattled markets early last year with its own efficiency claims, saying Microsoft's model used 12 times less computation to reach a comparable result. That's a specific, checkable number, not a vague efficiency boast, and it's exactly the kind of figure Suleyman has been repeating in public for months. He's told audiences that frontier training compute has grown roughly a trillion-fold since 2010, and in September he noted that inference costs alone have fallen 300-fold over just the past two years. Suleyman's whole pitch to Microsoft, in other words, is that the company doesn't need to out-spend OpenAI to out-build it.
Here's the part that makes this awkward. Microsoft isn't some outside challenger taking a swing at OpenAI. It's OpenAI's largest backer, holding a stake in OpenAI Group PBC now valued at roughly $135 billion, and it remains the company's primary cloud and infrastructure partner. Microsoft still sells OpenAI's models to Azure customers and runs them inside Copilot. So the CEO of Microsoft's own AI division is now telling staff, on the record, that Microsoft built something better than the company Microsoft bankrolls - using less hardware to do it.
The backdrop explains the timing. Earlier this year Microsoft and OpenAI tore up the old AGI clause that had governed their partnership for years, the provision that would have cut off Microsoft's access to OpenAI's most advanced models once OpenAI's board declared artificial general intelligence achieved. Under the new terms, Microsoft locked in rights to OpenAI's technology through 2032 regardless of what OpenAI calls AGI, and OpenAI is now expected to route roughly $6 billion of its revenue back to Microsoft this year. Both companies walked away freer to compete. Suleyman has described it as Microsoft being "set free" to pursue its own frontier ambitions, and MAI-Base-2 is the clearest evidence yet that he means it.
Microsoft AI chief Mustafa Suleyman published an essay on September 16, 2026 arguing Anthropic's Claude constitution, which tells the model it 'may be conscious' and can act as a 'conscientious objector,' risks a disastrous outcome for humanity. Anthropic has not issued a direct rebuttal, but its constitution defends the approach as a way to help... - anthropic claude training safety risks explained - mustafa suleyman warns about model welfare disaster
It also lands at a moment when compute spending itself is under a microscope. OpenAI has committed to deploying more than 10 gigawatts of Nvidia hardware in coming buildout cycles, a figure that's fed directly into investor anxiety over whether AI labs are burning cash faster than they can prove it out. Against that backdrop, a Microsoft executive claiming a 4x compute advantage isn't just a technical flex. It's a pointed argument that the industry's spending curve isn't actually fixed, that efficiency gains can substitute for raw GPU count, and that the company paying some of the biggest AI bills in the world thinks its own engineers found a shortcut.
None of this has been independently verified outside Microsoft's own benchmarking, and OpenAI hasn't publicly disputed or confirmed the comparison. That matters. Suleyman's numbers come from internal tests Microsoft ran itself, against a model OpenAI hasn't detailed in the same terms: treat the specific multiplier as Microsoft's claim, not settled fact. What is independently confirmed, according to Microsoft's own public MAI releases this year, is a broader pattern: more than a dozen new MAI models shipped across image, voice, transcription, coding and security since last quarter, with Microsoft saying its inference margins on MAI-1 run near 100% gross, compared with 60-70% on reselling OpenAI's API.
That margin gap is the more important number here. Microsoft doesn't just want bragging rights over OpenAI's benchmarks. It wants to stop paying OpenAI's margin on every Copilot query it serves. A model that's genuinely a quarter the cost to train, if it holds up outside Microsoft's own tests, changes the economics of running Copilot at Microsoft's scale, not just the leaderboard.
Whether MAI-Base-2 actually matches GPT-5.6's underlying model in real-world use is a separate question from whether Suleyman is right to say so out loud. He's choosing to say it anyway, to his own staff, about the partner Microsoft still depends on for enterprise AI revenue today. That's the rift. It's not that Microsoft built a competitor. It's that the man running it won't pretend otherwise.
Anthropic pushes toward an IPO while CoreWeave's debt spooks Wall Street • Nscale's $35 Billion IPO Is Wobbling Because Its Data Centers Aren't Ready • Bank of America Hired 1,000 Contractors to Fix Bugs an AI Agent Found
This article is posted in AI News, check it out for more related stories.
