Microsoft AI CEO: AI Improving AI Is ‘Dangerous’; 40% Pitch

Microsoft AI chief executive Mustafa Suleyman said AI models are now “fully involved in the process of improving their own models” and called that “definitely a dangerous capability”. He spoke on Yahoo Finance’s Deirdre Bosa Live in an interview recorded on Tuesday, 17 November, the day Microsoft launched six MAI models. The host posted the full interview on X late that evening, US time.

Our number: the cost line Suleyman gave in the same interview, “98% of the performance of GPT 6.1” Sol “at 40% of the cost”, is the figure Microsoft’s own footnote ties to a 50% promotional price. At full price the share is about 80%, by our arithmetic, as we reported on Tuesday.

Business Pill · Usage-Based Pricing

An evergreen Business Engineer explainer on usage-based pricing, paying for work rather than seats. It teaches the general idea only and does not discuss Microsoft or Suleyman.

The key insight: Microsoft’s AI chief called models improving their own models a dangerous capability while selling MAI on a 40% cost share that Microsoft’s own footnote ties to a promotional price, about 80% at full price by our arithmetic.

What Suleyman Said About Self-Improvement

Asked how recursive self-improvement fits Microsoft’s push to the frontier, he said “it’s pretty obvious that we’re starting to see models improve models”. Inside Microsoft, he said, models are used “to check for bugs, to produce training data, to create evaluations”.

He added: “I’m really, really nervous about it”. He said “the last three months have been a total watershed moment” and that “I don’t think that folks in the industry are being alarmist”. His stated approach: “proceeding step by step when we can first prove that they can be safe before we put them into production”.

MAI Code 2’s cost as a share of GPT-6.1 Sol’s: 40% as Mustafa Suleyman stated it on Yahoo Finance
MAI Code 2’s cost as a share of GPT-6.1 Sol’s: 40% as Mustafa Suleyman stated it on Yahoo Finance on 17 November, and about 80% at full price, because Microsoft’s own footnote ties the 40% to a 50% promotional price. Sources: Yahoo Finance, Deirdre Bosa Live; Microsoft AI, MAI Code 2 post and footnote, both 17 November 2026. The ~80% is our arithmetic.

What He Said About Microsoft’s Frontier Model

Suleyman said Microsoft is also building “full scale frontier models”. Asked whether they will compete with the best models from Anthropic and OpenAI, he said “in the new year, we’re hoping that they’ll get to the absolute frontier”.

On compute, he said: “compute isn’t a bottleneck for us at the moment”, and “We have gigawatts of compute across the Microsoft fleet”. He said Microsoft started its superintelligence team “12 months ago” and that a renegotiated contract with OpenAI “freed us up to go and pursue the absolute frontier”.

On open-weight models he said “I’ve always been a bit more skeptical of open source”, and that if powerful cyber capabilities can be run at home, “we’re gonna have a massive challenge of containment”.

How It Lines Up With Tuesday’s Data

OpenAI’s GPT-6.2 system card, released the same day, rated GPT-6.2 Astra its first publicly released model at High for AI self-improvement. On NanoGPT it reached 59% of the best human result, by our arithmetic. Epoch AI measured 12.3 serial reasoning steps per token without chain of thought, 17% more than GPT-6 Astra by our arithmetic, and said such models may be harder to monitor; Epoch called the figures preliminary.

OpenAI’s chief research officer, Mark Chen, put recursive self-improvement “somewhere between, let’s say, 25% and 50% there” on the same day, as we reported.

The Business Lever

The lever is price per unit of capability. Suleyman pitched MAI on cost, and the 40% he quoted holds only while Microsoft’s promotional price runs, per Microsoft’s footnote. He said the same company is pursuing the frontier on its own compute.

The Structural Read

As we read the interview, Suleyman described two tracks at once: cheaper specialised models now, and a general frontier model he hopes reaches the absolute frontier in the new year.

As we read the self-improvement remarks, he described models already doing parts of Microsoft’s own development, checking bugs, producing training data and creating evaluations, without naming a model or a benchmark.

Mustafa Suleyman, Microsoft AI, Yahoo Finance Deirdre Bosa Live, 17 November 2026

“I think that we do need to be extremely careful”

Three Implications

FOR MICROSOFT Suleyman said compute is not a bottleneck and that Microsoft has gigawatts of compute across its fleet.

FOR BUYERS OF MAI MODELS Microsoft’s footnote ties the 40% cost share to a 50% promotional price, about 80% at full price by our arithmetic, from our earlier reporting.

FOR THE SELF-IMPROVEMENT DEBATE OpenAI rated GPT-6.2 Astra High for AI self-improvement the same day, and Epoch AI said such models may be harder to monitor, from our earlier reporting.

The Business Engineer Lens

This story maps onto the Business Engineer framework Tokenomics: The Economics of AI.

The framework says of the token: “It is simultaneously the unit of cognition the model produces, the unit of compute the data center serves, the unit of price the lab charges, and the unit of value the enterprise extracts.”

As we read it, Suleyman pitched MAI on the unit of price the lab charges, while describing models that increasingly produce the work behind their own successors. The framework predates this interview and does not discuss Microsoft or Suleyman.

What Is Not Established

Suleyman did not name a model, a benchmark or a date for the watershed he described, and did not say how much of Microsoft’s model development is now done by models. He gave no release date for the frontier model beyond the new year.

Quotations are from our machine transcription of the interview video, checked against a second pass of the excerpt. The excerpt reached us through a post by @_Investinq. We did not contact Microsoft.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

Tokenomics: The Economics of AI

A Business Engineer analysis of the token as the unit of cognition, compute, price and value across the AI stack.

Read the Map of AI →

The Bottom Line

Microsoft’s AI chief called models improving models “definitely a dangerous capability” on the day he launched six cheaper models, quoting a 40% cost share that Microsoft’s own footnote ties to a promotional price, about 80% at full price by our arithmetic.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Quotations are from our machine transcription of Yahoo Finance’s Deirdre Bosa Live interview with Mustafa Suleyman, recorded 17 November 2026 and posted in full by the host on X; the on-screen branding identifies the show. The 98%, 40% and promotional-price footnote are from Microsoft’s MAI Code 2 post, as in our earlier coverage; the ~80% is our arithmetic. Epoch AI and OpenAI system-card figures are from our earlier coverage. We did not contact Microsoft.

Sources: Deirdre Bosa on X, full Yahoo Finance Deirdre Bosa Live interview with Mustafa Suleyman, 02:01 UTC, 18 Nov 2026 · @_Investinq on X (excerpt), 03:55 UTC, 18 Nov 2026

Scroll to Top
FourWeekMBA