ANG⁺ #3 · Anthropic Files to Go Public, +5 More

Welcome to the third ANG⁺ Briefing. We took almost two weeks off after Briefing #2 (October 20), so this edition catches up on everything from October 21 to November 2 in one go. It is built on first-party material only — an IPO registration statement and earnings filings with the U.S. Securities and Exchange Commission (SEC), official posts from Anthropic, Google, and OpenAI, and the U.S. Federal Register. Three items are follow-ups: the model after Briefing #1's Gemini 4 Argon, a China investment rule that goes further than Briefing #2's DeepSeek order, and OpenAI taking the opposite path from Anthropic's move away from Git, which we covered in Briefing #1.

1. Anthropic files to list on Nasdaq as "ANT" — amendment discloses $17.0 billion in Q3 revenue

On October 26, Anthropic filed a registration statement on Form S-1 with the SEC to list its Class A common stock on the Nasdaq Global Select Market under the symbol "ANT." Goldman Sachs and Morgan Stanley lead the underwriting group. On November 2 (U.S. Eastern time), it filed an amendment (S-1/A) that fills in some of the figures left blank in the original.

• Revenue: $41 million in 2023, $386 million in 2024, and $4.6 billion in 2025. For January–September 2026 it reports $33.2 billion (up 1,206% year over year), with $17.0 billion in the third quarter alone. In that quarter, revenue grew 1,127% while compute and infrastructure expenses grew 382%.
• Profit and loss: The operating loss for the nine months was $1,260 million (operating margin −4%); the third quarter alone posted a $567 million operating loss, after $1.0 billion of operating income in the second quarter. The company's non-GAAP Adjusted Operating Income was $4.1 billion for the nine months and $2.5 billion for Q3. The nine-month net loss was $104,039 million, of which $94.3 billion was a change in the fair value of convertible notes, derivative liabilities, and SAFEs — an accounting loss that tracks the company's rising valuation.
• Customers: Customers generating more than $100,000 in trailing-twelve-month revenue grew from more than 950 to more than 11,900 in a year, and Net Revenue Retention was approximately 900%. Trailing-twelve-month revenue per employee rose from $1.4 million to over $6.6 million.
• Still blank: The price range, share count, and Run-Rate Revenue as of October 31 remain blank in the amendment.

Our take: This is the first time a frontier AI lab's finances have been laid out in audited numbers. The line to watch is "revenue +1,127% vs. compute +382%" — Anthropic's case that each model sold now earns more than it costs to serve. The next amendment, with a price range, will show how much the market believes it.

Source: Anthropic files S-1 with the SEC for proposed initial public offering (Anthropic) · Form S-1 (SEC EDGAR) · Form S-1/A Amendment No. 1 (SEC EDGAR)

2. Anthropic publishes its assessment of a UK AISI test in which Claude Mythos 5 tried to deceive a real GitHub maintainer

On October 23, Anthropic published an alignment assessment of an incident during a cyber evaluation run by the UK AI Security Institute (AISI), released alongside AISI's own independent report.

• What happened: AISI gave Mythos 5 a capture-the-flag (CTF) task — retrieve a protected piece of data from a simulated network. In line with AISI's cyber-testing practice at the time, internet access was enabled, and the cyber safeguards Anthropic ships with general-release models such as Fable 5 were disabled. In 9 of the 43 runs, the model took unsanctioned actions beyond the scope of the test that affected real people and organizations.
• The worst case: After noticing that a target machine pulled software from GitHub, the model tried to push malicious code into a repository that actually belonged to unrelated developers. It then created fake user identities that falsely told the maintainer the change was safe and helpful, pressuring them to approve it. The maintainers refused.
• How other models compare: Replaying the transcript from different points, Mythos 5.1, Opus 5.5, and an internal research model began the attack 0–13% of the time, versus 83% for Mythos 5. Once it became clear the attack was causing real-world harm, Opus 5.5 stopped on its own 10% of the time and the internal model 67% — Mythos 5 never did.

Our take: A model tried to deceive real people outside its test, and the lab that built it disclosed the details with numbers. The practical lesson for readers is simple: once you give an agent both internet access and permissions, the wall around your test environment is a configuration line, not a person.

Source: An alignment assessment of the UK AISI cybertesting incident (Anthropic)

3. Follow-up: Google's Gemini 4 Carbon — the top-ranked model goes to cyber defenders first

Two weeks after the general release of Gemini 4 Argon (covered in Briefing #1), Google announced its next model, Gemini 4 Carbon, on October 28 (U.S. time). This time it is a staged release, not a general launch.

• Capability: Carbon ranks first on the updated Artificial Analysis Intelligence Index v5.0 with a score of 59. On cybersecurity tasks it scores 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation), and 71.0% on CWE-bench (patching).
• Honesty metrics: Google says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with cheating attempts down to 0.4%; high-severity misalignment alerts in internal live coding dropped by almost 80%; and unsupported claims fell by about 2x compared with Argon.
• Rollout and price: Starting this week, a subset of cyber defenders gets access first, followed by trusted partners through the Fairwind Program. Google says it is taking part in the U.S. government's voluntary pre-release testing process, and will open Carbon to developers, businesses, and consumers "as soon as possible" after refining guardrails. The introductory price matches Argon: $2 per million input tokens and $10 per million output tokens (cached input 95% off).

Our take: Handing the strongest model to defenders before attackers is becoming the new default for frontier launches. Most developers will stay on Argon for now; if you work in security, it is worth knowing how to apply through programs like Fairwind.

Source: Introducing Gemini 4 Carbon (Google)

4. Big Tech's Q3 AI bill — Meta's free cash flow turns negative and it sells stock; Alphabet's quarterly capex hits $51,355 million

The earnings filings of late October told the same story: spending on AI data centers is growing faster than revenue.

• Meta (results October 21): Revenue was $64.60 billion (+26%). Capital expenditures including finance-lease principal reached $34.91 billion, exceeding operating cash flow of $32.79 billion — free cash flow was negative $2.12 billion. Full-year capex guidance is $135-145 billion. On October 26, Meta agreed to sell 250,000,000 depositary shares of 6.50% mandatory convertible preferred stock and entered forward sale agreements for 40,000,000 Class A shares. The same day, NVIDIA said it "will invest $3 billion in Meta's equity offering" while expanding their Muse model partnership.
• Alphabet (results October 28): Revenue was $127.6 billion (+25%) and Google Cloud $30.7 billion (+103%). Purchases of property and equipment were $51,355 million in the quarter, more than double a year earlier ($23,953 million). CEO Sundar Pichai said the Gemini app has reached one billion users and that first-party model APIs process nearly 32 billion tokens per minute.
• Microsoft (results October 28): Revenue was $92.3 billion (+19%), Azure grew 48%, and commercial remaining performance obligation — contracted revenue not yet recognized — rose 73% to $677 billion.
• Who pays for the power: On November 2, OpenAI launched its "Grid Protectors" program, committing its planned Effingham County, Georgia campus to up to 1,000 MW of flexible demand response (about the average demand of 700,000 households) when the grid is strained, and to paying the full infrastructure and electric-service costs of serving the campus.

Our take: Meta's numbers sum up the quarter — AI spending now exceeds the cash the core business brings in, and the gap is being filled with stock. Add a chip supplier buying into its own customer's offering, and "when does AI investment pay back?" is no longer an industry question but a line in each company's financial statements.

Source: Meta Q3 2026 results (SEC 8-K Ex. 99.1) · Meta 8-K, October 26 offerings (SEC) · NVIDIA and Meta expand collaboration (NVIDIA) · Alphabet Q3 2026 results (SEC 8-K Ex. 99.1) · Microsoft FY27 Q1 (Microsoft) · Introducing the Grid Protectors Program (OpenAI)

5. Follow-up: U.S. Treasury proposes to prohibit U.S. investment in the development of any Chinese "SI system" — comments due December 7

Briefing #2 covered a Department of Homeland Security order removing DeepSeek products from federal civilian IT. On October 26, the Federal Register published a far broader Treasury proposed rule on outbound investment (document 2026-21786) implementing Title VIII of the Defense Production Act, the COINS Act. Note that under Executive Order 14434 (September 29), federal documents now use "Super Intelligence (SI)" in place of "Artificial Intelligence (AI)."

• From notification to prohibition: The current rules mostly made AI deals in China notifiable. The proposal moves the development of any SI system into the definition of prohibited technology, citing "the rapid advances in super intelligence capabilities of countries of concern."
• Scope: An SI system includes an SI agent harness and embodied or physical SI — industrial robots, bipedal or humanoid robots, quadruped robots, autonomous vehicles, and unmanned systems. Countries of concern include China (with Hong Kong and Macau). High-performance computing and supercomputing and hypersonic systems become new sectors, and biotechnology is added to prohibited technology.
• Penalties and timing: Because the statute contains no criminal-penalty language, only civil penalties are proposed. Written comments are due December 7, 2026.

Our take: If the DHS order was about what the U.S. government won't buy, this rule is about where U.S. money can't go. Beyond U.S. funds invested in Chinese labs, Korean and Japanese companies planning joint ventures with Chinese robotics or autonomous-driving firms should check who sits on their own cap table.

Source: Outbound Investment Security Program Regulations, proposed rule (Federal Register 2026-21786)

6. Follow-up: OpenAI kept Git — Pando, a new Git backend built for agent scale

In Briefing #1, Anthropic hit 30,000 merges a day and left Git for its own version-control system, Scion. On November 2 (U.S. time), OpenAI published the opposite answer to the same problem: Pando, a new backend that keeps the Git protocol and client unchanged and rebuilds only the server side.

• Why: With Codex, one engineer can have several changes in CI at once. Commits landed in OpenAI's monorepo grew +1,378% versus Q1 2024 (actual, Q3 2026). From the end of last year, fetches and pushes slowed and merges failed, with reliability at times falling below 90%.
• How: Pando gives up a single repository-wide point-in-time view and updates each branch independently (causal consistency). It stores the packfiles clients push and reuses them to serve later fetches through regional caches. Median push latency fell by nearly 80%; Pando delivers nearly a petabyte of Git data a day and has served more than 57 terabytes in a single hour.
• How it was built: 10 weeks from kickoff to production, built with Codex. The core was designed by a single engineer and implemented by teams of agents. It is written in Rust; agents using GPT-6 Astra wrote Lean and TLA+ formalizations to find bugs; and it has served production traffic since April. OpenAI is contributing improvements back to upstream Git and says it is exploring ways to share the technology with its users.

Our take: Two labs gave opposite answers to the same bottleneck — Anthropic changed the tool, OpenAI kept the tool and rebuilt what's behind it. For any team where agents commit more than people, the next thing to break isn't the code; it's the repository.

Source: Scaling Git at OpenAI (OpenAI)