Anthropic Files for an IPO With $16.3 Billion in Half-Year Revenue as OpenAI Ships GPT-6.1 Terra and Says Its AI Disproved a Complexity Conjecture

Anthropic Files Its S-1, Showing $16.3 Billion in First-Half Revenue and a Founder-Controlled Vote

The numbers that investors have guessed at for two years are now public. Anthropic’s Form S-1 registration statement on SEC EDGAR, filed Monday, asks to list Class A shares on the Nasdaq Global Select Market under the ticker “ANT.” Goldman Sachs and Morgan Stanley head a list of 22 underwriters. The price range, share count and run-rate revenue are still blank. The filing shows revenue of $41 million in 2023, $386 million in 2024 and $4.6 billion in 2025, then $16.3 billion in the first six months of 2026, which Anthropic puts at 1,300% year-over-year growth. As of September 30, more than 11,900 customers paid Anthropic $100,000 or more over the trailing twelve months, more than 1,900 paid $1 million or more, and about 250 paid $10 million or more. Anthropic says over half of the Fortune 500 are in that first group.

The headline loss is enormous, but most of it isn’t cash. Anthropic reports a net loss of $80.7 billion for the first half of 2026. Of that, $72.4 billion is the change in fair value of convertible notes, derivative liabilities and SAFEs, which the filing ties mainly to the company’s rising valuation. On its own adjusted measure, Anthropic swung to $2.7 billion of adjusted net income in the half, from an adjusted loss of $1.7 billion a year earlier. Compute and infrastructure cost $8.9 billion in the half, or 55% of revenue, down from 201% a year earlier. Control is spelled out in detail. A single Class T share lets the Long-Term Benefit Trust elect a majority of the board, and a single Class F share held by a founders’ LLC carries 50.1% of the general vote. Employees get a limited-voting Class E, and strategic partners hold non-voting Class N stock, which makes Anthropic a “controlled company” under Nasdaq rules. Reuters’ report on the Anthropic chip financing, via Investing.com, calls the listing potentially record-breaking.

“We do not expect our revenue will continue to grow at comparable rates, and we cannot assure you that it will continue to grow at all.” — Anthropic, Form S-1

Banks and Blackstone Line Up $63 Billion to Buy Broadcom Chips That Anthropic Will Lease

On the same day, Wall Street showed how Anthropic will pay for its next wave of compute. Broadcom’s investor release on the $63 billion financing structure says Bank of America, Blackstone, Citi and Morgan Stanley will fund a newly formed special purpose vehicle. The vehicle buys Broadcom XPU accelerators and networking gear and leases them to Anthropic. Bank of America, Citi and Morgan Stanley are arranging the first-lien debt, and Blackstone is the sole provider of the committed second-lien portion. Broadcom gives the vehicle residual value support equal to the first-lien amount, which is effectively a promise to cover the lenders if the hardware loses value. Deliveries are expected to start in calendar 2027 under a compute agreement Broadcom announced earlier. Reuters, citing two sources, puts the first-lien piece at about $43.5 billion, or 69% of the total. It notes that Apollo and Blackstone had already agreed to finance a $35 billion Anthropic expansion under an earlier Broadcom deal, with a Broadcom backstop of up to $29 billion, and that Moody’s warned the risk would rise if several such deals pile up in a short period.

“Fundamentally, the need for AI compute continues to grow at an extraordinary pace. This validates that there is capital available to finance such growth and that Broadcom is uniquely positioned to enable it.” — Hock Tan, President and CEO, Broadcom

OpenAI Launches GPT-6.1 Terra, a Cheaper Computer-Use Model at $1 per Million Input Tokens

OpenAI’s new mid-tier model is aimed at the screen work that fills an office day. OpenAI’s GPT-6.1 Terra launch post describes a fast model that combines visual reasoning, computer use and clearer writing. It is available now to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers as gpt-6.1-terra at $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. On Agents’ Last Exam, a benchmark of professional computer tasks, OpenAI says Terra at medium reasoning effort scores 15 points higher than GPT-5.6 Terra at about a third of the cost and matches GPT-6 Astra at a tenth of the price per token. At maximum effort, OpenAI estimates a cost per task about nine times lower than Anthropic’s Claude Opus 5.5. On its internal set of error-inducing prompts, the error rate at low effort fell from 53.6% to 16.7%. Terra solves 50.3% of Terminal-Bench 4.0 tasks at max effort, up from 21.5%, though OpenAI says GPT-6.1 Sol and GPT-6 Astra remain better for the hardest coding. The company also says Terra meets the Critical threshold in cybersecurity under its Preparedness Framework and keeps the safeguards built for Astra.

“Its standard API prices are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol, making it practical for developers to automate high-volume work, like reading incoming invoices and entering their details into accounting software.” — OpenAI

OpenAI Says Its Models Disproved the Exponential Time Hypothesis and Hands the Proofs to Outside Experts

OpenAI also made its boldest science claim yet. OpenAI’s post on working with academic communities announces three results produced by its internal models. The first is a proof of Sogge’s local smoothing conjecture in every dimension, which OpenAI says also resolves the Kakeya and Stein restriction conjectures. The second is a claim that the Exponential Time Hypothesis, a widely used assumption about how hard Boolean satisfiability problems are, is false. The third is an existence proof for quantum field theories including quantum chromodynamics. OpenAI says the ETH disproof was found on October 16 and formalized in the Lean proof assistant the next day. The math and computer science outputs are on GitHub. A group led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego, two of the hypothesis’s original proposers, and Ryan Williams of MIT will examine the ETH claim. Nobel laureate David Gross will lead the physics group, which is holding back the AI-produced material while it writes the work up. OpenAI says it seeks no academic credit and will fund a conference on the ETH result.

The Verge’s report on OpenAI’s three claimed discoveries adds an important caveat. The day before, OpenAI’s own panel of math advisers had written that “we recommend that AI companies do not adopt” the practice of choosing who sees results first, citing unequal access and borrowed reputation. All three results are still awaiting independent verification.

“Finding such algorithms, and turning theoretical advances into practical improvements, will require significant further research.” — OpenAI

Mirendil Debuts Fern, a Model Built Around a 10-Million-Token Context Window

A new lab is betting that context length matters as much as model size. Mirendil’s introduction of its first model, Fern, published Monday at 8:00 a.m. Pacific, says Fern was pretrained from scratch and supports up to 10 million tokens of context. That is roughly 23 MB of code, or about half a million lines, which Mirendil compares to the whole Megatron-LM library or days of an agent’s work. Fern 0.1 is the company’s base model after a short round of post-training, with 0.1 to 0.5 trillion parameters. Mirendil aims it at the work of people building AI: auditing training data for duplicates and benchmark contamination, debugging training runs from their logs and reading agent traces end to end. Access is listed as “coming soon” through an early-access form, and no price or API terms are published. Two larger base models are listed as more than 80% and more than 20% of the way through pretraining. The post says it tested next-token prediction and long-context access with the RULER benchmark, and that more comprehensive long-context results are coming soon.

“A model with a larger context window can use more task-specific information: draw on more of its own past experience, keep track of more specifications and instructions, integrate evidence from more documents and more code, and look at more data directly.” — Mirendil

Nevis Raises $70 Million to Put AI Agents to Work for Financial Advisors

Agents are moving into regulated back offices, and investors are paying up. Nevis CEO Mark Swan’s Series B announcement says the New York company raised $70 million led by EQT Ventures, with Charles Schwab, Sequoia Capital, ICONIQ, QuantumLight and Ribbit Capital, bringing total funding to $110 million in two years. Nevis combines CRM, meeting intelligence, financial planning and custodial workflows in one platform, with agents that handle tasks such as opening client accounts. The company says assets on the platform grew from $50 billion to more than $700 billion over the past year, it now adds $75 billion a month, revenue grew more than 25 times, and its agents have completed over 100,000 operational tasks. Customers include Focus Financial Partners, IEQ Capital and United Capital. Fortune’s profile of the former Revolut executives behind Nevis reports that the 70-person company was valued at $500 million, according to a person familiar with the matter. It also reports that Nevis has about 20 customers and opened a second office in London.

“There was an opportunity to give financial advisors access to AI agents that can actually do that manual work for them, so that they can save time, grow their book of business, and expand the level of service.” — Mark Swan, Nevis CEO, via Fortune