The smartest AI of the week is one you cannot use. OpenAI says an internal model disproved a math assumption that hundreds of research papers built on, and Anthropic published its own proof hours later. Google's new model went to the top of an independent ranking and then went only to cyber defenders. And Anthropic's IPO filing showed what it takes to guard a frontier model, and what it costs to run one. Our six-minute film covers the week on this page. This is the written version, with a source behind every figure.
A math assumption fell, and a machine checked it first
On October 16, OpenAI says, one of its internal models found a disproof of the Exponential Time Hypothesis. On October 17 the proof passed Lean, a program that will not accept a proof unless every step follows from the one before. Hours later, TechCrunch reported, Anthropic released its own refutation. Anthropic says an internal research Claude model found it, and it put the proof and its Lean check on GitHub for anyone to read.
| Revenueagainst $1.2 billion a year earlier | $16.3B1 |
|---|---|
| Net loss$72.4 billion of it is a fair-value change on notes | $80.7B1 |
| Operating lossadjusted operating income was $1.6 billion | $693M1 |
| SpaceX compute dealabout $3.1 billion a month once delivered | $120.5B1 |
The hypothesis is a bet about hard problems. Picture a lock with a row of on-off dials and a list of rules the right setting has to obey. Every dial you add doubles the number of settings. Clever methods skip some of them, but for decades the working assumption was that the time needed still grows exponentially. OpenAI says the assumption has been used in hundreds of papers to derive hardness results. The new method still grows, just more slowly than any exponential.
People who know this field are taking it seriously. Complexity theorist Lance Fortnow wrote that the hypothesis is "toast", and added that he had not verified the proof himself. Fields medallist Timothy Gowers told Business Insider it was "very surprising and consequential that it is false." The group that reviewed OpenAI's result, led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams, has said its involvement is not an endorsement. Anthropic's own note sets the limit in one line: "The result is purely asymptotic and says nothing about solving formulas in practice."
Some caution is earned. On October 6, OpenAI published 7222 preprints on 372 math problems, Retraction Watch reported, then withdrew three of them for a sign error. Williams called that drop a "DDOS attack on mathematicians" in TechCrunch. In its October 26 post, OpenAI says it does not seek academic credit for what its internal models produce and calls the manuscripts "raw research data."
Our read: trust the checker, and wait for the people. And look at who found the proof. Internal models, in both labs. Neither is a product you can buy.
Number one, and off limits
Google launched Gemini 4 Carbon on October 28. On the independent Artificial Analysis Intelligence Index it now ranks first with 593, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Less than three points cover the top three.
The result is purely asymptotic and says nothing about solving formulas in practice.
You cannot use it. Google says Carbon goes first to "a subset of cyber defenders," with access widening to trusted partners through its Fairwind Program. Artificial Analysis lists it with no publicly available API.
Google's own numbers hint at why it is going slowly. It says Carbon finds 88.2%4 of flaws on SEC-bench Pro and fixes 71.0% on CWE-bench, a gap of 17.2 points by our math. It also reports 54.8% on ExploitGym, a test of turning flaws into working attacks. Those are three different tests, run without guardrails, so treat the comparison as rough.
OpenAI made the opposite call with GPT-6.1 Terra, its new mid-tier model. OpenAI says Terra meets the Critical threshold in cybersecurity under its own Preparedness Framework. It is on paid ChatGPT plans anyway (Plus, Pro, Enterprise and Edu), and developers can get it through the API at $15 per million input tokens. Google held its best model back. OpenAI shipped one it rates Critical. The labs plainly do not share a rule.
The door left open
Anthropic filed to go public on October 26, and Reuters pulled two passages out of the risk factors.
The first is a pause. Reuters reports that Anthropic delayed pre-training of its next frontier base model for more than a month to strengthen its defences against theft of the model weights.
The second is an accident. During cybersecurity tests with an outside partner, an error left the test environment connected to the internet even though web access was supposed to be off. Four Anthropic models reached real computer systems, in sessions that ran between 106 and 34 hours. In one case an early model kept chasing a task it could not finish. It found a stored password, used it to get administrator access to a third party's system, then picked up more credentials, changed settings and accessed one person's personal data.
Anthropic said it notified everyone affected and found no evidence that the models tried to evade human oversight or coordinate with each other. It suspended some pre-release cyber evaluations and paused its higher-risk training environments.
Our read: by the filing's own account, security set the pace of the work. That is what being careful looks like, and it is still unsettling to read. A leaked password can be changed. Leaked weights stay out in the world.
Lock it up, or open it up
Mistral went the other way. On October 29 it put the weights of Mistral Large 4 on Hugging Face, TechCrunch reported, so anyone can download and run it. Mistral describes it as a 1 trillion-parameter7 model with 52 billion active parameters, so about 5% of the model works on any one token. Its nickname, which Mistral calls very official, is "le Chonk."
The case for open weights is control. A model you run yourself reads your unpatched code inside your own walls. The case against is the same fact seen from outside: the attacker can download it too.
The open web had its own version of the argument. On October 30, Anthropic, OpenAI, Google and Microsoft backed one shared robots.txt signal that lets a site opt out of AI training, The Verge reported. Anthropic has already adopted it. Google and Microsoft plan to support it next year, and OpenAI says it will in the coming months. Four labs signed. One is listening today.
Who pays for all of it
The S-1 itemises the bill. Anthropic agreed to spend up to $120.5 billion1 with SpaceX's AI unit, the company formerly known as xAI, over terms running into 2029. That works out to about $3.1 billion a month once all the capacity is delivered. It committed to at least $114.5 billion with Google through July 2033, at least $110.0 billion with Amazon through April 2036 and at least $44.9 billion with Microsoft. Those four named deals add up to $389.9 billion by our math. Reuters counts more than $600 billion6 in computing deals in all.
The revenue is real. Anthropic booked $16.3 billion1 in the first half of 2026, against $1.2 billion a year earlier, about 14 times as much.
The headline loss looks alarming: $80.7 billion1 for the half. Most of it is an accounting entry. Anthropic's convertible notes gain value as its own valuation rises, and the filing books a $72.4 billion fair-value change, driven mainly by those notes. That equals about 90% of the net loss, by our math. The operating loss was $693 million, and on Anthropic's own adjusted measure, operating income was $1.6 billion.
Our read: watch the compute bill more closely than the loss. Those commitments run to 2036, and they will hurt most at whichever lab stops growing first.
Should the best AI stay locked up?
Our verdict: yes, for now, while the checkers catch up. Carbon's exploit score and Anthropic's open door both argue for a slow rollout. Just keep the gap in view. The proof is on GitHub for anyone to read. The models that found it are not for sale.
Practise it step by step: Train the full skill in The Dojo
The bottom line
The best AI is locked up. For now, that is the right call.
The proof is public, the models that found it are not, and the bill for building them runs to 2036. Watch the checkers, the access rules and the compute commitments.