The Signal — October 27, 2026

Anthropic's IPO filing puts most of the control over the company in the hands of its seven co-founders and an independent trust. It also says Anthropic's own models are already contributing to the pace of its research. OpenAI says its models disproved a conjecture that hundreds of computer science papers rely on, and it has asked two of the people who proposed that conjecture to lead the checking. OpenAI also released GPT-6.1 Terra, a cheaper model built for computer use.

Anthropic's IPO filing gives its founders 50.1% of the vote through a single share, and a trust picks most of the board

Anthropic filed a public registration statement, known as a Form S-1, with the Securities and Exchange Commission on October 26 and plans to list on the Nasdaq Global Select Market under the ticker ANT, according to its announcement. It hasn't set the number of shares or a price range, and the filing isn't effective yet. The AP reports the filing sets up a Wall Street debut next month. NPR notes that the $1.5 trillion to $2 trillion valuations in circulation come from earlier press reports, not from the filing.

The filing lays out an unusual voting structure. Just before the offering, a company owned by the seven co-founders, which the filing calls the Founder LLC, will receive a single share of Class F stock. That share carries 50.1% of the total vote on ordinary shareholder matters, including the election of directors not chosen by the trust, and has no economic rights. Anthropic's Long-Term Benefit Trust, described in the filing as a body of "financially disinterested trustees with expertise in AI safety, national security, public policy, and social enterprise," holds Class T stock that lets it elect and remove directors who make up a majority of the board. Employees other than the co-founders receive Class E stock that votes only on limited matters. The founders decide how to vote the Class F share among themselves, one vote each. The share can be redeemed for nothing if they choose, if their LLC dissolves, or at the end of a transition period that starts once fewer than three "continuing founders" have remained for 90 consecutive days.

TechCrunch calls the structure controversial because voting power isn't tied to economic ownership, which leaves public shareholders with "little practical control" over the company. The filing reports revenue of $16.3 billion for the first half of 2026, compared with $4.6 billion for all of 2025, and says Anthropic doesn't expect growth at that rate to continue. Its operating loss narrowed to $693 million from $2.1 billion a year earlier. The $80.7 billion net loss for the half came mostly from a $72.4 billion noncash charge for changes in the value of convertible notes and similar instruments. As of September 30, Anthropic counted about 250 customers that each accounted for $10 million or more of its revenue over the previous twelve months.

Among the risk factors, Anthropic writes: "We have observed that our models are contributing to the pace of our own research." It warns that catastrophic or existential risks will grow if AI systems "accelerate AI progress beyond our ability to evaluate and safeguard new models." POLITICO reports that the filing also warns that the company's policy positions have drawn criticism from "influential political figures" and could bring investigations or adverse treatment in government procurement and visa decisions. NPR notes that the Pentagon designated Anthropic a "supply chain risk" this year after the company refused to let its AI be used for autonomous weapons or mass surveillance of Americans, and that Anthropic's two lawsuits against the Pentagon are still under way. Risk factors list what could go wrong rather than forecasting it, and nothing in the filing shows how the trust and the founders will use their votes once the company is public.

Sources: SEC (Anthropic S-1) · Anthropic · TechCrunch · POLITICO · NPR

OpenAI says its models disproved the Exponential Time Hypothesis, and researchers who proposed it will lead the checking

OpenAI announced three results it says its internal models produced: a refutation of the Exponential Time Hypothesis (ETH) in computer science; a proof of the local smoothing conjecture in every dimension, which would also settle the Kakeya conjecture in all dimensions; and a rigorous proof that quantum chromodynamics, the theory of the strong force, exists as a mathematical object. That physics result includes pure Yang-Mills theory, part of one of the Clay Mathematics Institute's Millennium Prize problems. No outside expert has confirmed any of the three yet.

ETH is a conjecture about how hard it is to solve Boolean satisfiability, the problem of deciding whether a logical formula can be made true. Researchers have widely believed it and have used it in hundreds of papers to argue that problems in graph theory, computational biology, databases and other fields can't be solved much faster. According to The Verge, the model's write-up describes an algorithm that counts the satisfying assignments of fixed-width formulas in subexponential time, and the case of 3-SAT is what would refute ETH. OpenAI says the model found the disproof on October 16 and that it was checked the next day in Lean, a language in which a computer verifies each step of a proof. The Lean proof and write-up are on GitHub. OpenAI says turning the result into faster practical algorithms "will require significant further research."

Instead of only posting the output, OpenAI asked outside experts to "shepherd" the results. The ETH group is led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego, who first proposed ETH with Francis Zane, and Ryan Williams of MIT. "Our involvement should not be taken as certification of the claim. A result of this significance requires careful, independent scrutiny," the three wrote in an open letter, according to The Next Web. Williams said the group signed no nondisclosure agreements and isn't being paid, and OpenAI agreed to the group's requests to publish all raw data and fund a conference. The physics group, led by David Gross, has chosen not to release the AI-produced material for now and will write up the work itself. The math result was posted directly, as OpenAI did with the 722 math manuscripts it released on October 6.

A Lean check shows that the formal statement holds as it was written down; experts still have to confirm that the statement matches the claim and that the setup is sound. The release process is also contested. The Verge reports that, the day before OpenAI's announcement, a panel of math advisers the company had assembled said AI companies should not give selected people first access to results, citing unequal access, borrowed reputation and biased selection. OpenAI's post acknowledges that choosing shepherds "necessarily shapes who gets involved first and creates asymmetry of information," and says it does not seek academic credit for the results.

Sources: OpenAI · The Verge · The Next Web · GitHub (Lean scope)

OpenAI's GPT-6.1 Terra is a cheaper model aimed at computer use

OpenAI released GPT-6.1 Terra. Terra is the second-cheapest of OpenAI's four model classes, which run in price order from Luna through Terra and Sol to Astra, according to 9to5Mac. It's available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, in the API, and in GitHub Copilot, where the rollout is gradual. API prices are $1 per million input tokens and $5 per million output tokens, half the standard prices of GPT-5.6 Terra and GPT-6.1 Sol.

OpenAI pitches Terra for office work done on a computer, such as interpreting technical drawings, filling out forms and editing videos. On Agents' Last Exam, a benchmark of professional tasks like turning architectural drawings into a 3D model, OpenAI says Terra at medium reasoning effort scores 15 percentage points higher than GPT-5.6 Terra at about a third of the cost, and matches GPT-6 Astra at a tenth of its price per token. On an internal set of prompts designed to provoke errors, OpenAI says the error rate at low effort fell from 53.6% to 16.7%. On Terminal-Bench 4.0, which tests coding agents working from the command line, Terra solved 50.3% of tasks at maximum effort against 21.5% for its predecessor. OpenAI says Sol and Astra remain better for the most complex coding. All of these figures come from OpenAI's own testing and haven't been reproduced elsewhere.

Simon Willison compares Terra's slot to Anthropic's Sonnet, "not the cheapest or the most expensive, making it harder to decide when to use it," and quotes OpenAI's Ted Sanders on Hacker News: "For subscribers, it's not a bombshell release." GitHub says that in early testing Terra solved noticeably more terminal and C# coding tasks than GPT-5.6 Terra. On safety, OpenAI's system card addendum treats Terra as Critical in cybersecurity and High for biological and chemical capability under its Preparedness Framework, so it ships with the same safeguards as the larger models. The addendum also reports a statistically significant regression on hate and extremism tests that pair text with images, compared with GPT-5.6 Terra; OpenAI says manual review found the model in line with prior launches.

Sources: OpenAI · OpenAI system card addendum · Simon Willison · GitHub Changelog · 9to5Mac

On the Editor's Desk

We held Atlassian's developer survey, which found that about a quarter of developers now say AI saves them ten or more hours a week, down from 68% last year. The numbers are self-reported, the report came out about a week ago, and Atlassian and InfoQ give slightly different figures (25% and 26%). Microsoft's buyout offer to about 60,000 employees describes who is eligible, not how many will leave, and nothing we found connects it to AI. A lawsuit accusing Anthropic of infringing Elvis Presley copyrights reached us only through a live blog and a copy of the complaint on a third-party site, so we're waiting for the court record. We also held OpenAI's new data on how people use ChatGPT for work, because its announcement and its download page cite different versions of the dataset.