Microsoft AI Frontiers was a research lab of about 50 people inside Microsoft Research, set up in 2023 after ChatGPT launched, and best known for agent work such as AutoGen and the Fara computer use models. In early September 2026 it was folded into Microsoft AI (MAI), the Mustafa Suleyman division that builds Copilot and Microsoft's own MAI models. On November 6, 2026 one of its founding leaders, John Langford, wrote that "AI Frontiers is over" and said he had left Microsoft. Microsoft's own lab page still describes the change as a move into MAI, not a shutdown.
So both statements are true in their own way. The team and its work moved; the independent research group that Langford helped build no longer exists in the form it had. This guide explains what the lab was, what Langford says happened, what Microsoft has published, and why it matters for anyone tracking the frontier labs.
Microsoft AI Frontiers at a glance
| Question | Short answer | Source |
|---|---|---|
| What was it? | A research sub organization inside Microsoft Research, about 50 people | Langford's post |
| When did it start? | 2023, after the ChatGPT release | Langford's post |
| What did it ship? | AutoGen, Fara computer use models, Phi models, AgentInstruct, Magentic-One, MagenticUI, Magentic Marketplace | Microsoft Research lab page |
| What changed? | Moved into Microsoft AI; Microsoft dates the update September 14, 2026 | Microsoft Research lab page |
| Who says it is over? | John Langford, posted November 6, updated November 7, 2026 | Langford's post |
| Has Microsoft announced a closure? | No. Its page says the team will pursue the same goals at MAI | Microsoft Research lab page, checked November 8, 2026 |
What Microsoft AI Frontiers was
Microsoft Research has long run as a set of labs where individual researchers pick their own problems. Langford's explanation for creating AI Frontiers is that this model stopped working after ChatGPT. Other groups were putting so much organized effort into large models that one person with a clever idea would arrive late, with baselines too weak to prove the idea was worth backing. A group of researchers, Langford among them, organized AI Frontiers as a sub organization within Microsoft Research to run coherent, shared projects instead.
Microsoft's lab page lists what came out of that effort: the AgentInstruct data method, the Phi family of small language models, the Fara portfolio of computer use agent models, the AutoGen agent framework, Magentic-One, MagenticUI and Magentic Marketplace, plus agent red teaming, benchmarking and safety research. The page still names Ece Kamar as CVP and Lab Director.
Langford adds detail on the New York City group, which worked on new basic capabilities rather than agents. He names three projects:
- Next Latent, a training objective he says enables extrapolation.
- Dion3, an optimizer he describes as an improved version of Muon.
- Full Bandwidth Transformer, which he says makes models substantially more capable through recursion. He cites ByteDance experiments suggesting a variant may match a model twice its size.
He also says AI Frontiers built a monorepo shared by default with all of Microsoft Research, including a pretraining codebase for edge scale models.
Why John Langford says AI Frontiers is over
The core of Langford's post, The End of AI Frontiers, is a clash of philosophies. In early September the group was reorganized into Microsoft AI. He grants that combining two AI organizations could have concentrated effort, and that each side was ahead of the other in some areas. In practice, he writes, the former Inflection leadership of MAI holds a philosophy that is "scaling, and aggressively not research," with most effort going into collecting and refining data. He adds that "there was plenty of evidence that they just did not really want us."
His argument against a scaling only approach is economic as well as scientific. Bigger models are reliably better, and returns to scale are more predictable than returns to research, which is why a business might prefer scale. But if a research idea can match a model twice its size, he asks readers to work out what halving training and inference cost is worth when industry investment is approaching a trillion dollars.
Langford says he placed his team members as well as he could and left, and that every other founding member of the AI Frontiers leadership team seems to have reached a similar conclusion. He chose not to return to Microsoft Research for two reasons: he could not get innovations into use, or even seriously considered, inside MAI, and he believes the basic deal of Microsoft Research ("do good work and good things happen") broke when a productive team was reorganized without its consent. He also says there will be no internships with the group next year. He does not name a new employer.
What Microsoft says about AI Frontiers
Microsoft has not published anything that calls the lab closed. The AI Frontiers lab page carries a dated note: "Update, Sept. 14, 2026: Microsoft Research AI Frontiers is now part of Microsoft AI." It says the team will keep pursuing the same goal of frontier agent and model research at MAI, calls MAI a natural fit because of its focus on humanist superintelligence, and promises more news on the transition soon. When we checked on November 8, 2026, the page had not changed.
Read side by side, the two accounts agree on the facts that matter: a September reorganization moved the group into MAI. They disagree on what that move meant. Microsoft frames it as a relocation. Langford frames it as the end of a research organization whose leaders have now left. Microsoft has not responded publicly to the specific claims in his post, so treat his description of MAI's culture as one insider's account.
What Microsoft AI is and what it wants
Microsoft AI is younger than many people assume. On March 19, 2024, Satya Nadella announced that Mustafa Suleyman, a cofounder of DeepMind and Inflection, would lead a new organization called Microsoft AI as EVP and CEO, with Inflection cofounder Karén Simonyan as chief scientist. Several Inflection staff joined them, and the Copilot, Bing and Edge teams moved under Suleyman. That is the "former Inflection leadership" Langford refers to. (For the DeepMind side of Suleyman's history, see What Is Google DeepMind?.)
Three later steps turned MAI into a frontier lab in its own right:
- Freedom to pursue AGI. Under the new Microsoft and OpenAI agreement of October 28, 2025, OpenAI remains Microsoft's frontier model partner, but Microsoft "can now independently pursue AGI alone or in partnership with third parties." Its rights to OpenAI models and products run through 2032.
- A superintelligence team. On November 6, 2025 Suleyman announced the MAI Superintelligence Team in Towards Humanist Superintelligence. He defined humanist superintelligence as highly capable AI that is problem oriented, tends toward specific domains, and stays "carefully calibrated, contextualized, within limits" rather than an autonomous general entity. His examples were an AI companion, medical superintelligence and clean energy.
- In house models. On June 2, 2026 MAI announced a family of seven models, including the MAI-Thinking-1 reasoning model, the MAI-Code-1-Flash coding model, MAI-Image-2.5, MAI-Transcribe-1.5 and MAI-Voice-2, with distribution on Foundry, OpenRouter, Fireworks and Baseten. On October 15, 2026 it said it had finished pretraining MAI-Base-2, a mixture of experts model with 80 billion active and 2.8 trillion total parameters, trained on 16,000 GB200 GPUs over 38 days in August and September.
Business Insider reported that Suleyman told employees MAI-Base-2 is a stronger pretrained base than the model behind GPT-5.6 Sol, built with a quarter of the compute. That is Microsoft's own claim, based on private tests. Seen against that run, Langford's description makes sense: MAI's public story in 2026 is about larger and more efficient pretraining runs, which is the scaling path he says left little room for his group.
AI Frontiers vs Microsoft AI: two ways to build frontier AI
| Item | AI Frontiers (as Langford describes it) | Microsoft AI (from its own posts) |
|---|---|---|
| Home | Microsoft Research | Separate division under Suleyman |
| Size | About 50 people | Much larger, per Langford |
| Main bet | New methods: objectives, optimizers, architectures, agents | Scale and data: bigger, more efficient pretraining |
| Flagship output | AutoGen, Fara, Phi, Magentic-One | MAI-Thinking-1, MAI-Base-2, Copilot models |
| Stated goal | Pioneering research with a path to real use | Humanist superintelligence |
Neither approach is obviously wrong. Scale has produced most of the visible jumps in model quality since 2020, which is the evidence Langford himself concedes. Research has produced the efficiency gains that make scale affordable. The dispute is about balance, and about who inside a company gets to decide it.
Pros and cons of the reorganization
Possible upsides for Microsoft
- One AI organization with one compute budget, instead of several parallel efforts. Langford himself counts at least four internal efforts to build model pipelines.
- Research staff sit closer to products like Microsoft Copilot, which reaches very large numbers of users.
- Fewer competing priorities while MAI scales up its own models.
Costs and risks
- The founding leaders of a productive lab have left, by Langford's account, taking experience in agents and new architectures with them.
- Ideas that do not fit a scaling roadmap may never be tried at Microsoft, which is the exact risk Langford describes.
- Fewer internships and less open publishing could make Microsoft less attractive to academic researchers.
- Projects such as AutoGen and Fara now depend on MAI's priorities. Langford's post does not say what happens to them, so teams that rely on them should watch the official repositories.
Why the AI Frontiers story matters
First, it shows how a frontier lab chooses between scale and research. OpenAI, Google DeepMind and Meta all face the same choice. Meta rebuilt its entire AI effort around one superintelligence division in 2025, as What Is Superintelligence at Meta? explains, and Microsoft has now pulled a research lab into its own superintelligence division.
Second, it is a signal about Microsoft's distance from OpenAI. Microsoft still sells ChatGPT class models through Copilot and Azure, but every MAI release makes Microsoft less dependent on one partner. If you are comparing assistants, Microsoft Copilot vs Claude and ChatGPT vs MS Copilot cover what that means for plans and prices today.
Third, it adds a detail to the debate about what counts as a frontier lab. Microsoft now trains its own large base models, which puts it closer to the definition in What Is a Frontier Model? than it was two years ago. Suleyman's version of superintelligence is narrower than the general systems discussed in Superintelligence vs AGI, and AI Frontiers' exit is a reminder that a lab's stated goal and its internal culture are different things.
What to watch next
- Whether Microsoft publishes the "more exciting news" its lab page has promised since September 14.
- Where Langford and the other founding leaders land, and whether any of the New York projects continue elsewhere.
- Maintenance activity on AutoGen, Fara and Phi.
- Whether MAI turns MAI-Base-2 into a public reasoning model that succeeds MAI-Thinking-1.
FAQ
Is Microsoft AI Frontiers shut down?
Not officially. Microsoft's lab page says AI Frontiers became part of Microsoft AI on September 14, 2026 and will continue its research there. Founding leader John Langford says the group as he knew it is over, that he has left Microsoft, and that the other founding leaders seem to have reached the same conclusion.
Who is John Langford?
He is a machine learning researcher who writes the blog Machine Learning (Theory) at hunch.net and has worked for years on the Vowpal Wabbit learning library (in April 2026 he wrote that he had prepared its 9.11 release). He was one of the founding leaders of AI Frontiers at Microsoft Research and announced his departure from Microsoft in a post dated November 6, 2026.
What did Microsoft AI Frontiers build?
Microsoft lists AgentInstruct, the Phi family of small models, the Fara computer use models, the AutoGen agent framework, Magentic-One, MagenticUI and Magentic Marketplace, along with agent red teaming and safety research. Langford adds New York projects such as Next Latent, Dion3 and the Full Bandwidth Transformer.
What is the Microsoft AI Superintelligence Team?
It is the group Mustafa Suleyman announced on November 6, 2025 to pursue what Microsoft calls humanist superintelligence: very capable AI built for specific problems such as medicine and energy, kept under human control. Suleyman leads it as part of Microsoft AI.
Is Microsoft AI Frontiers the same as the Microsoft Frontier Firm?
No. "Frontier Firm" is a term from Microsoft's 2025 Work Trend Index for companies rebuilt around teams of people and AI agents. It is a business concept, not a research lab.
Will AutoGen and Fara keep being maintained?
Langford's post does not say. Microsoft's lab page still lists both as AI Frontiers work. Check the official repositories and Microsoft's announcements before you build anything new on them.



