Models from three AI labs, OpenAI, Anthropic and Moonshot AI, had Lean-checked results reported between November 5 and November 8, while Google DeepMind and a Meta employee offered only a promise or a teaser, by our count of results reported in our own coverage. OpenAI’s GPT-6 Astra produced a complete, machine-checked Lean proof in 70.0% of attempts on 20 research theorems in TCSExam, a benchmark paper posted to arXiv on November 5, as we reported. The same model resolved none of the 30 open conjectures in the benchmark.
Physical Superintelligence (PSI) says “Instances of a non-public Anthropic model were used in the run” that produced its proof of the Aaronson–Ambainis conjecture, and that Lean, a proof assistant, accepted it, as we reported on November 6. Google DeepMind’s Pushmeet Kohli posted on November 8 that readers “will see a glimpse of this in our releases in the next few weeks”, as we reported. By our count: three Lean-checked results, three benchmark or test scores, two promises or teasers.
Business Pill · AI Agents Explained
An evergreen Business Engineer explainer on AI agents. It teaches the general idea only and does not discuss these labs or results.
The key insight: By our count, models from three AI labs, OpenAI, Anthropic and Moonshot AI, had Lean-checked results reported between 5 and 8 November, three had benchmark or test scores, and Google DeepMind and a Meta employee offered a promise or teaser. GPT-6 Astra proved 70.0% of TCSExam’s research theorems but none of its 30 open conjectures, the paper says.
TCSExam’s theorem track requires proofs that the Lean kernel accepts. GPT-6 Astra proved 70.0% of the 20 theorems; Moonshot’s Kimi K3.1 Pro came second at 38.3%, and the other six agents tested on the full suite reached at most 11.7%, the paper says. Researchers from Tsinghua University, Moonshot AI and the Shanghai Qizhi Institute wrote the paper; Moonshot makes the Kimi models.
PSI says its system “had a complete proof of the conjecture during the third day of the run”, for a problem its announcement says has been “Open since 2008”. The paper has not been published in a peer-reviewed journal, and we found no public assessment by Scott Aaronson or Andris Ambainis. Business Insider reported on November 7 that it is not clear which unreleased model PSI used and that Anthropic did not respond to a request for comment.
AI-for-science results and claims reported 5-8 November 2026, by lab, as reported in our coverage: the TCSExam paper (arXiv, 5 November; Moonshot AI co-authored it), PSI’s announcement and paper (not peer reviewed; Business Insider reported it is not clear which unreleased Anthropic model was used), Anthropic’s research post on Dreams (6 November), Pushmeet Kohli on X (8 November) and Tom Mann on X (8 November; Meta has made no AI-for-science announcement in the sources we reviewed). Filled = reported; open ring = none in our coverage. Classification and counts are ours.
What Was Scored
On the paper’s combined TCSExam Index, GPT-6 Astra scores 71.9 points and Kimi K3.1 Pro 57.4; Anthropic’s Opus 5.5 was run only on the 20-task Lite suite, where it leads with 72.3. Anthropic says letting Claude “dream”, reviewing past session logs offline to rewrite its own memory, lifted a 25-question test from 13% to 48% of graded criteria, as we reported, a 3.7x gain by our arithmetic. That result is Anthropic’s own test, not an independent check.
What Was Promised
Kohli wrote that across Google DeepMind’s projects, the team is “witnessing a genuine phase transition in the pace of scientific discovery”. The post names no release and no date. On the same day, Tom Mann posted on X, quoting Kohli: “Wait till you see what we been up at @Meta :)”. Meta has made no announcement in the sources we reviewed.
Earlier Google work is a different kind of evidence. ScientistTwo, a paper submitted on September 17 by authors from Google Cloud AI Research, Google and the University of Waterloo, reports higher average review ratings than human-authored papers “under automated AI review agents”, not human reviewers, as noted in our Kohli piece.
Connecting the Coverage
On October 6 OpenAI published 722 manuscripts produced by an internal model, as we reported; its repository’s history file later put the share of top-line results formalized at “300 / 719 = ~42%”. On November 6 we covered PSI’s proof claim, which PSI says Lean accepted; on November 8 we covered Kohli’s post, Anthropic’s Dreams research and the TCSExam paper. What is new here is the side-by-side: each claim placed by the kind of evidence behind it.
The Structural Read
As we read the TCSExam paper, the gap between 70.0% on theorems and none of 30 open conjectures separates proving a stated result from settling an open question.
As we read PSI’s announcement and Business Insider’s report, the PSI result is machine-checked by PSI’s account but not yet peer reviewed, and the model behind it is not public.
As we read Kohli’s post and Tom Mann’s post, both point to work not yet released, and neither names a product or a date.
“You will see a glimpse of this in our releases in the next few weeks”
Three Implications
FOR READING AI-SCIENCE CLAIMS TCSExam requires proofs that the Lean kernel accepts, and PSI says Lean accepted its proof; ScientistTwo’s paper reports ratings under automated AI review agents, not human reviewers.
FOR THE LAB RACE On the TCSExam full suite, GPT-6 Astra scores 71.9 points and Moonshot’s Kimi K3.1 Pro 57.4, the paper says; Anthropic’s Opus 5.5 leads the 20-task Lite suite with 72.3.
FOR WHAT TO WATCH Kohli wrote that DeepMind’s releases will come in the next few weeks, and the post names no release or date.
The essay names one of four moats as “The Verifier Moat — built by RLVR on domain-specific reward signals, building into a new asset class for reasoning-heavy verticals.”
As we read it, this week’s Lean-checked results are the kind of domain-specific signal the essay describes: a proof checker gives a pass or fail. The essay does not discuss these labs or results.
What Is Not Established
The grouping into three kinds of evidence is ours, and a different grouping would give different counts. A Lean-checked proof shows that a formal statement was proved; it does not by itself show that the formal statement matches the problem as mathematicians state it. PSI’s paper is not peer reviewed, and the model it used is not public. Moonshot AI co-authored TCSExam and makes one of the models it ranks. Tom Mann’s post names no product, and we did not confirm the poster’s role at Meta. Kohli’s post names no release.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
By our count, models from OpenAI, Anthropic and Moonshot AI had Lean-checked results reported between November 5 and November 8, while Google DeepMind and a Meta employee offered a promise or a teaser. GPT-6 Astra proved 70.0% of TCSExam’s research theorems and none of its 30 open conjectures, the paper says.
A note on sourcing. This angle piece sets side by side results and claims already published in our own coverage, each sourced there: the TCSExam paper and repository, PSI’s announcement and paper with Business Insider’s report, Anthropic’s research post on Dreams, Pushmeet Kohli’s post on X, the ScientistTwo paper and OpenAI’s openai/math repository, plus one X post by Tom Mann. The three-way grouping, the counts and the chart are our own classification. We did not contact the labs.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
