Two of the most consequential events in AI infrastructure this week landed on the same day. Anthropic filed its S-1 with the SEC on 26 October, placing its financials on public record for the first time. OpenAI launched GPT-6.1 Terra the same day, halving API prices for agentic and coding workloads. In parallel, both labs published competing AI-generated proofs on an open problem in theoretical computer science, and the UK government committed £100 million to a permanent AI biosecurity laboratory. Each development would have dominated the week on its own; together they mark a distinct shift in what AI looks like as formal infrastructure.

Anthropic files its S-1: a first look at AI revenue at scale

On 26 October, Anthropic, PBC filed its registration statement with the US Securities and Exchange Commission for a proposed initial public offering on Nasdaq under the ticker ANT. The filing is a required public document and carries audited financial data, risk factors, and governance terms — none of which Anthropic had previously disclosed.

The revenue figures are the most consequential single data point the lab has ever made public. First-half 2026 revenue was $16.3 billion, against $1.2 billion in the same period a year earlier: roughly a 14-fold increase in twelve months. Eighty per cent of that revenue was consumption-based API usage; the remaining 20% came from subscriptions. International revenue accounted for $4.5 billion. Q2 operating income was $974 million, meaning Anthropic is profitable at the operating line. The headline net loss of $80.7 billion is almost entirely non-cash: $72.4 billion of it is an accounting charge on convertible notes and does not reflect cash consumed.

The enterprise cohort data matters for partners and operators. As of 30 September, Anthropic counted more than 11,900 customers with $100,000 or more in trailing twelve-month revenue, up from fewer than 950 a year earlier. Customers at $1 million or more numbered more than 1,900. Infrastructure commitments stood at approximately $265.6 billion as of June 30. The seven co-founders retain 50.1% of voting control through a single Class F share, which means governance authority remains with the founding team regardless of what happens in the public market.

The filing also discloses that reliance on a limited number of customers and distribution channels is a material risk, and that ongoing tensions with the Trump administration and the Department of Defense could cause "material revenue losses or business disruptions."

OpenAI launches GPT-6.1 Terra: API prices halved for agentic workloads

OpenAI released GPT-6.1 Terra on 26 October with standard API pricing set at $1 per million input tokens and $5 per million output tokens — half the rate of both its predecessor GPT-5.6 Terra and the recently released GPT-6.1 Sol. The model carries a context window of 1,050,000 tokens and a maximum output of 128,000 tokens. It is available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and in the Codex environment.

The model is positioned for coding, computer use, and agentic tasks. One pricing caveat warrants attention for teams running long-context pipelines: requests exceeding 272,000 input tokens are billed at double the standard input and cache rates, and 1.5 times the standard output rate. For most agentic workloads this threshold is unlikely to bind frequently, but multi-document retrieval and long-session agents should account for it in cost projections.

Independent benchmarks place Terra marginally below GPT-6.1 Sol on composite scores. The performance gap is small; the cost gap is 50%. For teams running high-volume agentic pipelines, the pricing change is likely more consequential than the benchmark difference. The effective economics of agentic deployment just changed materially.

Competing AI-generated proofs: the Exponential Time Hypothesis and Kakeya under review

OpenAI announced on 26 October that its internal models had produced three results in mathematics, theoretical computer science, and physics, releasing them to academic communities for review. The computer science result is a claimed refutation of the Exponential Time Hypothesis (ETH), which holds that 3-SAT cannot be solved in subexponential time. OpenAI says its model produced a deterministic algorithm counting satisfying assignments in subexponential time, formalized the result in Lean, and engaged the ETH's original authors as external reviewers.

The mathematics result is a claimed proof of the Kakeya conjecture (via the local smoothing conjecture) in every dimension. The physics result addresses Yang-Mills existence, one of the Clay Millennium Prize Problems. OpenAI released the CS and mathematics proofs publicly on GitHub; the physics result was written up by outside researchers rather than distributed directly.

On the same day, Anthropic released a parallel paper, attributed to Claude, on the same 3-SAT problem. Scientific American notes that the releases conflict with guidance from OpenAI's own Advisory Group on Mathematics and Artificial Intelligence, and that the prompts, compute time, and model version have not been disclosed. All results remain under independent review.

The operational read: if any of these results survive verification, they represent a qualitative shift in what AI systems can contribute to formal research. The fact that two independent labs produced competing claims on the same open problem in the same week — without coordination — is itself significant data about the current capability level of frontier models. Operators in research-intensive sectors, and anyone whose competitive position depends on algorithmic or mathematical work, should track the verification process over the coming weeks.

UK commits £100 million to an AI biosecurity laboratory

The UK government announced on 23 October the establishment of the AI Biosecurity Lab (AIBL), a £100 million independent non-profit funded by the public sector as part of the National AI Biosecurity Initiative. Its mandate is to develop defences against AI-enabled biological threats.

OpenAI and Anthropic are each contributing up to £100 million, primarily as model credits and access to frontier models. OpenAI's contribution is structured through its Rosalind Biodefense programme. Anthropic is providing model credits, model access, and technical support. Google DeepMind is contributing researcher time and Gemini model access, though the financial terms of its involvement have not been disclosed. The UK government is separately committing £65 million toward treatment and countermeasure development for dangerous pathogens.

The AIBL is the first institution in any jurisdiction to treat AI-enabled biosecurity as a permanent infrastructure problem. The model-credits structure is worth noting: the contributing labs are not making cash contributions but providing proprietary access at scale, which creates a dependency relationship between the lab and its funders. For operators in healthcare, defence, or regulated industries, the lab's work will likely shape both technical standards and procurement frameworks over the next several years.

The four items this week share a common thread: scale meeting formal accountability. Anthropic's S-1 closes the era of opaque AI financials and places enterprise contract terms in a new context — partners can now read the filing and assess what they represent in Anthropic's revenue base. The price cut on GPT-6.1 Terra makes large-scale agentic deployment economically viable for a wider range of organisations. The mathematical results, still contested, push into territory where AI output requires specialised human infrastructure just to evaluate. And the biosecurity lab creates a permanent institution between frontier capabilities and the governments expected to govern them. None of these is a headline that will fade; each is a structural condition that operators will be working within for years.