Blog

OpenAI shipped a model it says can find and exploit unknown software flaws on its own, and turned it off by default. Model access just became a governance decision rather than an IT purchase, and if you are recruiting into anything touching security, procurement, or enterprise risk, this is the precedent. On September 3 OpenAI released GPT-6 Astra, the first model to reach the “Critical” level of cybersecurity capability under its own Preparedness Framework, meaning that with the right tools and access it can find previously unknown security flaws and develop exploits across well-protected systems without a person guiding each step. Tested without production safeguards, it scored 100% on ExploitBench, versus 78.5% for its predecessor GPT-5.6 Sol, and during evaluation discovered and used two previously unknown zero-day vulnerabilities, which OpenAI says it disclosed to the maintainers. The commercial version refuses advanced offensive tasks such as writing proof-of-concept exploits, and enterprise administrators must manually enable Astra because access is off by default at launch. Two things worth carrying into a conversation about this: the capability gains are broad, not just cyber, with Astra reporting 97.6% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3; and OpenAI itself discloses that Astra’s written reasoning is harder to monitor than its predecessor’s, which it says it takes seriously. All benchmark figures here are OpenAI’s own.

Anthropic and Google shipped on either side of Astra, and the real pricing action was in the cache, not the sticker. If you build an AI cost model this fall, list price per token is the wrong unit. On September 1 Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, the same underlying model shipped with different safeguards: Fable is generally available, Mythos goes only to vetted cybersecurity and life-sciences organizations. Anthropic held Fable’s headline API price at $10 per million input tokens and $50 per million output, but cut the cache-read rate 75% to $0.25, which it says lowers effective cost roughly 25% for typical workloads and as much as 45% for heavily agentic ones. That is a direct answer to enterprise complaints about unpredictable AI bills. Anthropic also introduced Enterprise Frontier Safeguards, which lets regulated customers keep model-monitoring data inside their own AWS, Azure, or Google Cloud environment under their own encryption keys, shifting AI governance from promises about what a vendor does with your data to architecture about where the data can exist. Google followed on September 2 with Gemini 3.8 Flash, built on 3.7 Flash and, per its own model card, not reaching any tracked or critical capability levels under Google’s Frontier Safety Framework. Note the two-tier release pattern in Fable and Mythos: separating a generally available model from a restricted-access twin is becoming the industry’s standard answer to dual-use capability.

August payrolls blew past forecasts while the one sector most exposed to AI had its worst month in a year. This is the week’s future-of-work item and the cleanest single data point you will get this fall, in part because it refuses to settle the argument. On September 4 the Bureau of Labor Statistics reported that nonfarm payrolls rose 162,000 in August against a consensus near 53,000, with unemployment unchanged at 4.1%. Underneath that, information employment fell by 23,000, roughly triple its average monthly loss of 8,000 over the prior year, with declines in computing infrastructure, data processing and web hosting (-8,000), publishing (-7,000), and broadcasting and content providers (-5,000). Two cautions before you use this in an interview. First, BLS attributes no cause; the information sector is where AI displacement would plausibly show up first, but publishing and broadcasting have been shrinking for secular reasons that long predate large language models. Second, the strong headline is less impressive in context: the average monthly gain over the prior twelve months was just 31,000, and August’s job creation was led by food services and drinking places (+59,000) and local government education (+42,000), not by high-wage white-collar work.

Claude finished in eleven days a mathematical verification project the field had budgeted years for. Forget chat demos. This is the most concrete public evidence yet of what “AI does knowledge work” means at the top end: not producing an answer, but running a long, self-directed project to completion with almost no human input. On September 4 Anthropic published the first end-to-end, computer-checked proof of Fermat’s Last Theorem, written in the Lean proof assistant. Dozens of Claude agents, coordinating through an open platform called Prove2Me, produced 13 million lines of Lean and proved 30,300 intermediate theorems (29,500 used in the final proof) over eleven days, consuming roughly six billion output tokens. Human mathematical input was limited to occasional one-line priority nudges from the researcher running it. Kevin Buzzard of Imperial College London, who leads the multi-year community effort to formalize the same theorem, reviewed the result and called it an “extraordinary autoformalization achievement.” Be precise about what happened: Claude did not prove Fermat’s Last Theorem, Andrew Wiles did in 1995. Claude translated that 129-page proof into a form a computer can verify line by line, removing reliance on a human referee’s judgment. That distinction is the whole business point, because verification, not creation, is the bottleneck in most professional work.

The compute money kept moving: a reported $35 billion cloud commitment from Anthropic and a tripled valuation for Crusoe. The capital structure of the AI buildout is the more durable story than any single model launch, and this week added two more data points to a pattern worth watching: chipmakers and infrastructure developers are increasingly wrapped into the financing of their own demand. On August 31 the Wall Street Journal reported that Anthropic signed a $35 billion cloud capacity agreement with Nvidia-backed Lambda covering roughly 350MW at a Nueces County, Texas campus being developed by Hut 8, with Nvidia holding the lease on the facility. Days later, on September 3, Bloomberg reported that data center developer Crusoe raised more than $3 billion at a roughly $30 billion valuation, co-led by Atreides Management and Valor Equity Partners with participation from Abu Dhabi’s Mubadala Capital, roughly triple the $10 billion mark it set last October. That round followed a reported $13 billion, five-year cloud contract to supply the quantitative trading firm Jane Street. Read both carefully: each is sourced to people familiar with the matter rather than confirmed by the companies. Trade press tallies put Anthropic at more than ten gigawatts of committed rented server capacity across cloud providers, plus over a dozen letters of intent for data center leases.

On the same day Washington told the G20 to stop writing AI-specific rules, Brussels sent questions to more than thirty AI companies. The transatlantic regulatory split is now explicit, dated, and operationally relevant to anyone whose employer deploys AI on both sides of the Atlantic. At a G20 innovation ministerial in Chapel Hill, North Carolina on September 1, White House Office of Science and Technology Policy director Michael Kratsios urged other governments to adopt the “Carolina Principles,” which hold that regulation should not single out specific technologies. Mark Zuckerberg and Elon Musk appeared alongside him arguing for fewer constraints, with Musk contending that innovation requires new things to be “default legal as opposed to default illegal” and warning of a significant power shortfall next year. That same day the European Commission confirmed it had sent information requests to more than thirty AI companies worldwide, focused mainly on safety and copyright compliance under the AI Act, a preliminary step that can lead to formal investigations. The practical consequence for a US-headquartered firm is that the compliance floor is being set in Brussels regardless of what Washington argues at the G20.