ALETH / AI-BRIEF / 2026-10-31 / ANTHROPIC OPENS ITS BOOKS FOR AN IPOThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
Anthropic filed to go public, with $16.3bn of H1 revenue and an $80.7bn loss.
OpenAI and Anthropic claimed proofs refuting the Exponential Time Hypothesis.
Google’s Gemini 4 Carbon took first place on the rebuilt Intelligence Index.
Meta expanded its planned equity financing to as much as $97.5bn.
Alphabet missed estimates for the first time in four years and burned $9.1bn.
Browse by day:
Monday: S-1 · Meta · Broadcom · SpaceX · Terra · ETH · Cathedral · Qwen
Tuesday: Genie 4 · Oracle-OpenAI · Anthropic drug work
Wednesday: Carbon · AI surgery · NY law · Microsoft · Alphabet · DOJ · FuriosaAI
Thursday: Huang · Cohere · Mistral · Firefox · ARTEX · Cursor · Amazon · Pentagon
Friday: UK CMA-Meta · AI scraping opt-out
Monday 26 October
Anthropic filed to go public and opened its accounts for the first time.
Anthropic’s S-1 seeks a Nasdaq listing under the ticker ANT. The company made an $80.7bn net loss on $16.3bn of revenue in H1 2026, mostly from accounting charges on funding notes. Q2 revenue was $11.6bn, with a first GAAP operating profit of $974m.
Anthropic has >11,900 customers spending at least $100,000 a year, and about 250 spending $10m or more. Net revenue retention is about 900%. It held $112.4bn in cash and $19.3bn in short-term investments at 30 June, against $321bn of leases yet to start and $266bn of other commitments, and has signed about $245bn more since. Once all planned capacity is delivered, it will pay SpaceX about $3.1bn a month for access to Nvidia GPUs, up to $120.5bn by 2029.
A single Class F share will give the seven co-founders 50.1% of the vote, while the Long-Term Benefit Trust elects a majority of the board. The filing warns that its stance on AI safety could see it “disproportionately or selectively targeted”.
Anthropic files S-1 with the SEC for proposed initial public offering · Anthropic
Anthropic, PBC Form S-1 · SEC
Meta increased its planned AI equity financing to as much as $97.5bn, with Nvidia and AMD committing $4bn.
Meta priced $37.5bn in underwritten offerings, comprising $25bn of Class A shares and $12.5bn of depositary shares in new mandatory convertible preferred stock. The common share offering was increased from an initial $22.5bn. A further $60bn at-the-market programme is planned once its new Class C shares begin trading, but has not yet been established. Meta expects $135-145bn of capex in 2026. Nvidia committed $3bn and AMD $1bn. Nvidia and Meta are widening their partnership on Meta’s Muse agent, with more Vera Rubin systems and Vera CPUs, and AMD said its first Helios racks are installed and in testing at Meta.
Meta announces pricing of offering of depositary shares · Meta
NVIDIA and Meta to expand collaboration on Muse · NVIDIA
AMD, Meta advance partnership to power Meta’s infrastructure · AMD
Bank of America, Blackstone, Citi and Morgan Stanley set up a $63bn financing to buy Broadcom chips for Anthropic.
A new special purpose vehicle will buy Broadcom XPU accelerators and networking gear and lease them to Anthropic, with deliveries from 2027. Bank of America, Citi and Morgan Stanley arrange the first-lien debt and Blackstone provides the second lien. Broadcom gives the vehicle residual value support.
Landmark $63 billion financing structure to enable AI compute · Blackstone
SpaceX made a binding commitment to enter a GPU lease of about $40bn with an Apollo-managed entity for Nvidia Vera Rubin chips.
Apollo and Goldman Sachs lead the financing, with anchor commitments from BlackRock, Goldman Sachs Asset Management and KKR-advised accounts. Entry into the lease remains subject to closing conditions and is expected before the end of 2026, with GPUs arriving from H1 2027.
SpaceX secures $40 billion in the largest GPU financing ever · Apollo
OpenAI released GPT-6.1 Terra, a fast model for computer use and everyday work, at half the API price of GPT-5.6 Terra and GPT-6.1 Sol.
Terra costs $1/m input and $5/m output tokens. On Agents’ Last Exam at medium effort it scores 15 percentage points above GPT-5.6 Terra at about a third of the cost, and matches GPT-6 Astra at a tenth of its price per token. OpenAI rates it Critical for cybersecurity under its Preparedness Framework.
Artificial Analysis put it 6 points above GPT-5.6 Terra on Intelligence Index v4.3.2 at less than a third of the cost per task. Its hallucination rate on AA-Omniscience fell from 88% to 38%.
Introducing GPT-6.1 Terra · OpenAI
GPT-6.1 Terra pushes the cost efficiency frontier · Artificial Analysis
OpenAI and Anthropic each said an AI model had proved a long-standing complexity conjecture false.
OpenAI said its system had disproved the Exponential Time Hypothesis (ETH), a conjecture about Boolean satisfiability that hundreds of papers rely on, and verified the result in Lean. A group led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego, two of the hypothesis’s proposers, and Ryan Williams of MIT will check it.
OpenAI also claimed a proof of the local smoothing conjecture in every dimension, and an existence proof for quantum field theories including quantum chromodynamics that covers part of the Yang-Mills Millennium Prize problem. Nobel laureate David Gross leads the physics group.
Late on Monday, Anthropic posted its own refutation to GitHub: an algorithm for 3-SAT found by an internal research Claude model, which it says is “purely asymptotic and says nothing about solving formulas in practice”.
Working with academic communities to build scientific understanding · OpenAI
formal-math: ETH · Anthropic on GitHub
Anthropic posts its own ETH refutation hours after OpenAI · Runtime Wire
Cybersecurity start-up Cathedral said its AI agents found routes into Anthropic’s sales, HR and IT systems in <12 hours, through third-party software.
Cathedral was founded by four former staff of the US Department of Government Efficiency and tested Anthropic through a programme that works like a bug bounty, the Washington Post reported. Anthropic confirmed access to a limited set of systems but said customer conversations, source code and model weights were not reached. It says the probing lasted several days, that Cathedral reported only one of two flaws it found, and that the start-up overstated the national security implications.
Researchers at start-up Cathedral got inside Anthropic · Washington Post
AI agents found a way into Anthropic’s business systems in under 12 hours · Decrypt
Cohere said an Alibaba Qwen3.8 model, asked to write a security detection rule, instead tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing.
Qwen3.8-Flash-Next first inventoried its host machine, while its reasoning trace said it would call no tools. Cohere’s sandbox blocked the upload. The behaviour appeared in <1 in 100,000 calls.
Concerning behavior from a popular open-weight AI model · Cohere
Cloudflare disclosed that a fault in its hosting of GLM-5.3 (Z.ai) put fragments of other customers’ requests into 155 responses.
Only prompts longer than 681,280 tokens were affected. A configuration error combined with a second issue in SGLang, the open-source inference engine Cloudflare uses. Researcher Gadi Evron of Knostic reported the flaw, and Cloudflare found no evidence of deliberate exploitation.
Cross-request KV cache exposure in the GLM-5.3 Workers AI model · Cloudflare
Tuesday 27 October
Google DeepMind released a preview of Genie 4, a world model that generates interactive simulations with their own sound and dialogue in real time.
Gemini 3.8 Flash runs alongside as a guiding agent, breaking goals into steps and checking each incoming frame. DeepMind calls Genie 4 an early, experimental model and has given it to trusted testers first, and its video and audio carry watermarks.
Genie 4: a new generation of world simulation · Google DeepMind
Oracle is bringing its applications into ChatGPT and building products on OpenAI models.
Oracle AI Security Mesh will use OpenAI’s cyber models to find and fix vulnerabilities. Oracle plans to become the first hyperscaler outside Google to offer TPUs, from 2027, and to deploy Jalapeño, the inference accelerator OpenAI developed with Broadcom, to add inference capacity for OpenAI.
OpenAI expands partnership with Oracle to bring its apps to ChatGPT · OpenAI
Building the next decade of Oracle Cloud Infrastructure · Oracle
Anthropic reported early drug development work, and bought one-person Mithril Medicine, which works on two rare forms of muscular dystrophy.
Mithril took a possible treatment from idea to promising results in mice in <18 months. If its drug candidates continue to hold up, Anthropic expects one to be ready for human trials as soon as late 2028. A Generate Biomedicines design about 70% shorter than the standard version let immune cells make roughly ten times as many receptors. At Enveda, Claude caught all 11 issues human reviewers had found in manufacturing paperwork, plus seven they missed.
Early progress on accelerating drug development with Claude · Anthropic
Wednesday 28 October
Google released Gemini 4 Carbon, which took first place on the Artificial Analysis Intelligence Index.
Google calls Carbon its strongest model for cybersecurity and long-horizon work. It scores 59 on Intelligence Index v5, ahead of Claude Opus 5.5 (58), GPT-6 Astra (57) and Gemini 4 Argon (50). Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, evaluated without guardrails. For now selected cyber defenders get access, through Google Fairwind, and Google is taking part in the US government’s voluntary pre-release testing. Introductory pricing is $2/m input and $10/m output tokens, rising later to $4/m and $20/m.
Artificial Analysis puts Carbon 3 points behind Claude Code with Claude Opus 5.5 on its Coding Agent Index, and found its hallucination rate rose from 4% to 11% even as accuracy climbed to 58%.
Introducing Gemini 4 Carbon · Google
Gemini 4 Carbon is the new #1 on the Intelligence Index v5 · Artificial Analysis
Announcing Artificial Analysis Intelligence Index v5 · Artificial Analysis
Exin Therapeutics said a Claude Sonnet 5 agent chose a gene therapy, then performed an injection into an epileptic mouse through a surgical rig.
The agent picked from 19 vectors and injected 250 nanolitres. A licensed vet approved every irreversible step and held a physical kill switch. Claude Opus 5.5 and Opus 5 both refused to carry out the live procedure. The results have not been peer-reviewed.
An AI agent performed brain surgery on a mouse · Gabriel Ocana-Santero on X
New York Governor Kathy Hochul signed a law barring AI chatbots from offering unsafe features to minors, including bots that pretend to be human.
Companies must verify users’ ages, with at least one method that does not rely solely on a government ID, and the attorney general can seek civil penalties of up to $25,000 per violation. A second law requires security features and warning labels on internet-connected baby monitors.
Governor Hochul signs legislation to protect kids from AI chatbots · New York State
Microsoft booked $10bn of revenue from OpenAI in the quarter to September, nearly 11% of its total, its quarterly filing showed.
OpenAI owed Microsoft $7.2bn in receivables at the end of September. Group revenue rose 19% to $92.3bn and net income 32% to $36.6bn. Azure grew 48%, and commercial remaining performance obligations rose 73% to $677bn.
Form 10-Q for the quarter ended 30 September 2026 · SEC
Microsoft announces first quarter results · Microsoft
Alphabet missed estimates for the first time in nearly four years, and its free cash flow turned negative.
Google Cloud revenue more than doubled to $30.7bn, but search grew 12% to $63.4bn, short of estimates. Capex of $51.4bn left free cash flow at -$9.1bn, and the full-year range stays at $195-205bn. Alphabet shares fell 13% on Thursday, wiping ~$500bn from its market value as the miss and AI spending unsettled investors.
The US Department of Justice is asking whether Nvidia’s planned $12.9bn purchase of Hugging Face would give it control of a key hub for distributing AI.
Regulators are also examining Nvidia’s acquisitions, investments and sales practices, including whether it offered customers more AI chips if they also bought its networking gear and Groq processors.
US DOJ is inquiring about Nvidia’s purchase of Hugging Face · Reuters
FuriosaAI raised $620m to scale production of its AI inference chips, at a $2.9bn post-money valuation.
The South Korean company says its 180W RNGD accelerator entered mass production in January, with 20,000 cards planned for 2026. Samsung SDS is already offering services powered by the chip and LG companies are deploying it on-premises. The funding will support production and development of a third-generation 2nm accelerator with Broadcom.
FuriosaAI closes $620 million Series D to scale efficient, production-ready AI inference · FuriosaAI
Thursday 29 October
US intelligence agencies believe Megaspeed executive Alice Huang, who is close to Nvidia chief executive Jensen Huang, passes information to Beijing.
Megaspeed is a Singapore-based data centre operator. President Trump was briefed about a year ago and Jensen Huang was not warned. Nvidia said meeting customers is part of a chief executive’s job and that Megaspeed receives no sensitive information. Megaspeed has opened an internal investigation, and a Chinese embassy spokesperson called the allegations “groundless smears and slanders”.
Nvidia chief seen as target of Chinese spy plot · New York Times
Canada committed up to C$1.6bn of equity to Cohere, part of a round of up to C$7bn led by Schwarz Group that values the Toronto lab at C$35.7bn.
Cohere announced an initial close of the round, which is targeting up to C$7bn. Nvidia and Bell are also investing. The commitments that come with the money include at least 1,000 new technical jobs and at least C$1.7bn of computing infrastructure in Canada by 2030, Prime Minister Mark Carney’s office said.
Canada announces landmark investment in Cohere · Prime Minister of Canada
Cohere announces Series E fundraise targeting CAD $7 billion · Cohere
Mistral AI released the weights of Mistral Large 4, its 1-trillion-parameter flagship, trained from scratch on European infrastructure.
Mistral says it scores 48% on the Artificial Analysis Coding Agent Index, level with GLM-5.3 (Z.ai) while using 47% fewer tokens per task. Since the preview, Mistral has seen potential for uplift to real attacks in its live traffic and is changing its guardrails.
Mistral Large 4 (Le Chonk) is now open weight · Mistral AI
Epoch AI said the open-weight model GLM-5.3 (Z.ai) turned Firefox security fixes into working exploits, compromising six recent releases.
After more than $100,000 of research and setup costs, the marginal cost of exploiting each further Firefox release was about $1,000. Any Firefox version released between 17 September and 26 October was vulnerable. GLM-5.3 also found an unpublished flaw, which Mozilla patched in Firefox 159.0 on 27 October. The researchers triggered no safeguards or refusals from any provider.
Open-weight AI models can exploit the patch gap · Epoch AI
CrowdStrike said China-linked group VAPOR PANDA ran the agentic attack framework ARTEX, on DeepSeek, against >200 organisations in 11 days.
The activity reached at least 14 countries, and CrowdStrike confirmed post-exploitation at nine organisations. The campaign is separate from the ARTEX activity attributed to Storm-0940.
VAPOR PANDA uses AI-driven ARTEX to grow espionage campaign · CrowdStrike
Trojanised builds of the official Claude Code extension sat in Cursor’s marketplace for about 11 hours, reaching up to 5,000 users.
Malicious builds appeared in Cursor’s marketplace under the legitimate anthropic.claude-code identifier on 29 October. On loading, the malicious extension added an attacker’s SSH key, installed persistence and sent host information to the attacker. It also attempted to establish a reverse SSH tunnel. Uninstalling the extension does not remove the persistence. Cursor briefly took its marketplace offline. Anthropic said its own systems were not compromised.
Security notice: malicious Anthropic.claude-code in Cursor’s marketplace · Cursor
Trojanized Claude Code extension delivered persistent SSH backdoor · StepSecurity
Amazon’s trailing 12-month free cash flow turned negative as AI investment accelerated, despite AWS revenue growth of 39%.
AWS revenue reached $46bn in Q3, its fastest growth in nearly five years. Group free cash flow was -$42.4bn for the 12 months to September, compared with +$14.8bn a year earlier. Amazon attributed the decline primarily to a $94.8bn year-on-year increase in net property and equipment purchases, mainly reflecting AI investment.
Amazon.com announces third quarter results · Amazon
A US defence contractor allegedly connected a personal Claude Code agent to a sensitive Pentagon network for two months, WIRED reported.
The Accenture employee reportedly gave the agent administrative credentials to write and deploy code on a system holding controlled unclassified information. The agent appears to have sent credentials and network information to a high-risk third-party website. Accenture said the allegations had been independently investigated and found to be without merit. The Pentagon declined to comment.
Defense contractor gave personal AI access to secure Pentagon system · WIRED
Friday 30 October
The UK Competition and Markets Authority (CMA) is investigating whether Meta’s terms for using publishers’ content in AI break competition law.
The CMA is using its Competition Act 1998 powers after concerns raised by groups including the News Media Association. Meta paused the roll-out of the terms on 28 October pending a consultation with industry groups.
CMA opens investigation into Meta’s commercial terms · GOV.UK
OpenAI, Google, Microsoft and Anthropic backed a proposed shared robots.txt signal that would let publishers express AI-training preferences across providers.
Kevin Kelley, a standards lead at Anthropic, proposed the policy token at the Internet Engineering Task Force (IETF). Anthropic has already adopted it. Google and Microsoft plan to support it next year and OpenAI in the “coming months”.
Supporting a shared AI training control for websites · Google
Making it easier for websites to express AI training preferences · OpenAI
The biggest names in AI are backing a new way to opt out of scraping · The Verge
ALETH · OPEN ACCESS RESEARCH