On November 19, 2026, Business Insider reported that Sergei Gukov’s Caltech lab was building open-source AI models for mathematical problems. Separately, SAIR’s Open Math OS soft-launched on October 16, 2026, as a collaborative workspace with existing models; SAIR’s math-focused model was still in development in Terence Tao’s launch announcement. Business Insider’s report and Tao’s announcement describe distinct projects, not two names for the same system.

When you hear about a math-AI project, check whether it’s a model, a research workspace, or both.

What is Gukov’s Caltech lab building?

Gukov and a handful of students are reported to be building math-focused models from scratch, with most designed to solve one specific problem. A broader model was still in development and not ready at the time of the interview, Business Insider reported.

Gukov also told the outlet that some of his models had solved a few questions before OpenAI’s October 2026 release of hundreds of mathematical solutions. The account does not identify those questions; the report attributes the claim to Gukov.

What did Open Math OS host at its soft launch?

Open Math OS launched as a shared research workspace using existing models, while SAIR was assembling partners to post-train an existing open model for mathematics. Tao’s October 16 announcement named DeepSeek V4.1 Flash, Kimi K3, and GLM 5.3 Flash as the models hosted at that soft launch.

Tao described those models as roughly a year behind the frontier at the time. He said they could support work such as medium-sized Lean formalization and discussion of mathematical arguments, but were not suited to the one-shot problem-solving described in recent AI announcements. That makes the launch snapshot useful for understanding the workspace’s role: it brought tools and hosted models together before SAIR’s planned math-focused model was ready.

What can researchers do in the workspace?

Open Math OS combines shared projects and chats with AI-assisted problem exploration, computation, Lean proof formalization, and LaTeX writing. The project’s official FAQ lists more than 100 preinstalled math and scientific-computing tools, including Lean 4, Mathlib, SageMath, PARI/GP, SymPy, Python, and LaTeX.

Its Ultramath feature runs several AI agents on a problem in parallel, shares their findings, and returns proposed solutions for the user to assess, according to the FAQ. The distinction matters in practice: an agent’s proposed argument is something to examine, while a proof formalized in Lean can be machine-checked.

The FAQ also says users’ work is not used for model training by default; training use requires explicit consent and agreed terms. It says users retain ownership of their work and credit for their ideas.

What does a live Open Math OS session look like?

Watch Terence Tao demonstrate Open Math OS with Lean 4 and Ultramath on Wilson’s theorem

In a SAIR workshop demonstration, Terence Tao works through Wilson’s theorem in Open Math OS using Lean 4 and Ultramath agents. It is a concrete example of the collaborative workflow described by the project.

Experts say OpenAI’s final swarm took a different route from the Spanish mathematicians’ program