TabPFN-3.5 tops the tabular benchmarks with a thinking mode
Three stories about money and rules: the enterprise data model that just got a reasoning variant, the odd-couple summit asking Trump for a superintelligence treaty, and Washington's plan to make Mexico's AI hardware North American.
Prior Labs shipped TabPFN-3.5, a tabular foundation model it says now leads both of the external benchmarks the field uses for real-world business data. The company reports the base model finishing about 150 Elo points ahead of the previous overall leader on BeyondArena, and up to roughly 250 Elo points over the strongest prior baseline on the hardest slice — text-rich, high-cardinality and high-dimensional datasets where the columns are product codes, customer notes and free-text review fields. On its own benchmark pages, TabPFN-3.5 stands first on TabArena and first on BeyondArena, and Prior Labs says it holds SOTA at 1 million rows and up to 20,000 features.
The interesting part is the family, because it mirrors what the LLM labs did over the past two years. TabPFN-3.5-Thinking trades inference compute for accuracy — plus 44 Elo on TabArena and about 20 more on BeyondArena over the base model — and is hosted on the API. TabPFN-3.5-Plus adds heavier text processing for messy string columns and ranks second behind Thinking on STRABLE, a benchmark of 108 tabular datasets built around dirty strings. TabPFN-3.5-Fast is an alpha checkpoint the company says runs up to six times faster. The base model handles text and datetime columns natively, which matters more than any single score: the reason gradient-boosted trees have outlived four generations of transformer hype is that feature engineering is where enterprise time goes, and a model that eats raw ERP and CRM tables with no preprocessing attacks the labor, not just the accuracy.
Tabular prediction is the unglamorous half of enterprise AI, and it is the half with a budget attached — cash-flow forecasting, payment delay, supplier risk, churn. The distribution tells the same story: SAP completed its acquisition of Prior Labs in July after committing more than €1 billion, and TabPFN-3.5-Plus is available in SAP AI Core on day one, with AWS SageMaker and Microsoft Foundry integrations alongside it. Xiaomi open-sourced its own tabular foundation model two weeks ago — Xiaomi's open TabLDM takes aim at the last thing LLMs never fixed: tables — so the category now has an open challenger and an owned incumbent. The caveat: the arenas are community-run but the headline Elo deltas are vendor-reported, and "Thinking" for tabular data is a new cost profile nobody has priced yet.
Bernie Sanders used the main stage of the Future of Life Institute's Pro-Human Assembly to ask President Trump to bring an AI treaty to next week's summit with Xi Jinping. Speaking in Washington on Tuesday, the Vermont senator called for a global agreement including Beijing to ban superintelligence and pause advanced development, invoking Reagan and Gorbachev at Reykjavik: "With the survival of humanity at stake, they found a way to sign that treaty." He said he will introduce legislation next week to permanently ban artificial superintelligence and pause frontier work until safety rules exist, and called Congress "asleep at the wheel" and the president's ignorance on the subject "truly embarrassing."
The coherence problem showed up eight minutes later, when Steve Bannon spoke — on the same program, not the same stage — and reached for the identical cold-war analogy to argue the opposite. Bannon wants an economic and technological "quarantine" of China: Chinese nationals out of US research labs, an end to training Chinese engineers, capital flows severed, "no deals, no agreements." His only path to regulation runs through defeating China first. Both men denounced the same oligarchs for the same reason, and split on what to do about the one variable that actually constrains American policy.
What makes the assembly more than theater is the mix on the roster. Representative Greg Casar, co-sponsor of Sanders's bill, named the counter-lobby by fund — "AI billionaires are spending millions through the Leading the Future PAC" — and Republican Chip Roy rejected the premise that anyone in particular should decide: not the president, not the CEOs, not Congress alone. Fifty-plus organizations used the same day to launch a Coalition for Pro-Human AI. We covered the bill's introduction in September — Sanders and Casar move to ban superintelligence and pause frontier AI.
The United States is pressing Mexico to accept rules of origin for AI hardware that would cap how much of a chip or a server can come from outside North America, according to the Wall Street Journal. The proposal is aimed at Chinese firms routing around US tariffs rather than at Mexico itself, and it is being negotiated inside the talks to revise the US-Mexico-Canada Agreement — the trilateral deal Trump has repeatedly threatened to abandon. A new round of trade talks is expected in Washington as early as next week.
The number that explains the pressure: AI hardware overtook automobiles this year as Mexico's largest export to the United States. That makes component sourcing a lever with real reach — Mexico's maquila plants assemble and ship servers and boards that Chinese suppliers feed, and a rules-of-origin threshold converts a tariff dodge into a paperwork failure at the border. It also puts Mexico in the position it has occupied for a year: Washington's preferred manufacturing address and the customs chokepoint for everyone else's.
For AI buyers the effect is slower and more expensive than any single tariff. Rules of origin do not tax a shipment; they force a supply chain to be rebuilt and re-documented to prove where the parts came from, and the cost lands in the price of servers already constrained by memory and packaging. Watch the timing: if Washington holds the USMCA revision until Mexico accepts the sourcing rules, the AI hardware map gets redrawn as a trade concession rather than an industrial policy.
What to watch: whether the House Energy and Commerce Committee moves the FRONTIER Act before the fall recess, and whether the Sanders bill reaches markup with a treaty ask attached to it.
If the biggest AI budget line is a spreadsheet model that never writes a sentence, has the industry been arguing about the wrong frontier? Tell us in the comments.
Sources: TabPFN (GitHub)
