1. Google’s Gemini 4 Carbon: its strongest model, but only for cyber defenders — for now

On October 28, Google DeepMind announced Gemini 4 Carbon, two weeks after Gemini 4 Argon. Google calls it “our strongest model with state-of-the-art cybersecurity and long horizon capabilities.” The numbers it leads with are security numbers: 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for validating them through exploitation, and 71.0% on CWE-bench for patching them. Google notes those security scores were measured without guardrails. On internal tests, Carbon more than doubled Argon’s success rate at generating secure code, and Google says its teams have seen it work for days on complex engineering problems.

Carbon also took the top spot on the newly revised Artificial Analysis Intelligence Index v5, scoring 59, one point ahead of Claude Opus 5.5 at 58.

The catch is access. Carbon goes first to “a subset of cyber defenders” through Google’s Fairwind Program, widening gradually to trusted partners. Google has set no date for developers, businesses, or consumers, and no general price beyond the introductory rate it says will match Argon’s. In other words, the week’s top-ranked model is, for most people, a press release with a waitlist attached.

Why it matters

This is a different kind of frontier launch: capability framed as defence, not as a chatbot upgrade. A model that finds, validates, and patches flaws in one loop is genuinely useful to security teams — and, in the wrong hands, the same loop is an offensive tool. Google’s staged rollout is the honest response to that dual use. The question to watch is whether “defenders first” stays a sequencing choice or becomes the template for how the most capable models are distributed.

Good reads

2. Anthropic files its S-1: $16.3B in six months, and a very large asterisk

On October 26, Anthropic publicly filed its Form S-1 for a proposed Nasdaq listing of Class A shares under the ticker ANT. The share count and price range are blank, the registration is not yet effective, and no shares can be sold until the SEC signs off. Goldman Sachs and Morgan Stanley are listed first among the underwriters. Reports put a possible listing in the week of November 16, potentially raising $100 billion or more at a $1.5–2 trillion valuation — those figures are press reports, not in the filing.

The filing gives public investors their first audited look at a frontier lab’s finances, and the headline numbers are extreme in both directions. Revenue for the six months ended June 30, 2026 was $16.286 billion, up from $1.163 billion a year earlier and more than triple all of 2025 ($4.585 billion). The second quarter alone produced $11.6 billion in revenue and $974 million in GAAP operating income. The half-year operating loss narrowed to $693 million.

Then the asterisk: a $80.748 billion net loss for the same six months. About $72.4 billion of it is a non-cash revaluation of convertible notes, derivative liabilities, and SAFEs as Anthropic’s valuation rose — accounting, not cash burn. Anthropic held $131.651 billion in cash, cash equivalents, and short-term investments at June 30. Usage-based fees made up 80% of first-half revenue, subscriptions the rest.

Why it matters

This is the first major AI lab to file publicly, and it reframes the funding debate with real numbers. The growth is real and steep; so is the compute bill behind it. A public listing will force the question the private market has deferred: whether a company can keep growing revenue 14x year over year while its largest cost — compute — grows nearly as fast. Watch the amended filings for price, size, and any Q3 numbers before treating the reported valuation as a fact.

Good reads

3. OpenAI’s GPT-6.1 Terra: computer use at half the price

Also on October 26, OpenAI released GPT-6.1 Terra, its mid-tier model, replacing GPT-5.6 Terra. It is built for the mundane work that fills an office day: drafting proposals, reading technical drawings, filling out forms, editing video, reading invoices and entering them into accounting software. It is available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and through the API as gpt-6.1-terra.

The price is the story: $1 per million input tokens, $0.05 cached input, $5 per million output — half the price of GPT-5.6 Terra and GPT-6.1 Sol. OpenAI’s own benchmarks claim the model scores 15 percentage points higher than its predecessor on Agents’ Last Exam at medium effort, at about a third of the cost, and matches GPT-6 Astra on that benchmark at a tenth of the price per token, completing tasks more than three times faster. On Terminal-Bench 4.0 it solves 50.3% of tasks at maximum effort, up from 21.5%. Its low-effort factual error rate fell from 53.6% to 16.7%, per OpenAI. All of these are vendor numbers; independent tests are not out yet.

Why it matters

Frontier capability keeps getting cheaper, but Terra is a different kind of price cut: it targets the tier where agents actually do volume work. If the cost-per-task claims hold up even roughly, computer-use automation stops being a demo budget item and becomes something a team can run all day. That moves the bottleneck from model price to controls — what the agent is allowed to click, submit, and pay for — which is exactly where this week’s oversight letters (see below) are pointing.

Good reads

Elsewhere this week

  • Google DeepMind’s Genie 4: worlds that talk back (Oct 27): a real-time world model with native audio — ambient sound, music, and dialogue generated alongside the visuals — and live steering by typed prompt while the scene runs. Coherent scenes last up to about two minutes, double Genie 3’s limit, and details still drift on longer runs. Research preview for trusted testers, with Google AI Ultra subscribers in Project Genie and Gemini API partners next; all output carries SynthID watermarking. (Silicon Report, Eesel)
  • Artificial Analysis Intelligence Index v5 rewrites the scoreboard (Oct 27): 60% of the index now rests on private test sets or held-out answers, with two new evaluations, FrontierCode 2.0 and Terminal-Bench-Science. At launch Claude Opus 5.5 led at 58, ahead of GPT-6 Astra at 57; Carbon’s 59 a day later took the top spot. The uncomfortable footnote: Kimi K3.1 Pro scores 46, but its weights are still unreleased — among models with downloadable weights, DeepSeek V4.1 Flash and GLM-5.3 lead at just 32, 26 points behind the leader. A v5 score is not comparable to a v4.3 score. (CellCog, Crypto Briefing)
  • House Democrats ask frontier labs to report incidents within 48 hours (Oct 27–28): ranking members Pallone, Lofgren, and Raskin wrote to Anthropic, Google, Meta, OpenAI, and SpaceXSI requesting a full accounting of misalignment and safety incidents by November 11, notice within 48 hours of any new incident, and within 24 hours if critical infrastructure appears involved. This is a request from minority-party ranking members, not a law or a subpoena — but it sketches the reporting regime Congress may legislate next. (House Science Committee Democrats, FourWeekMBA)
  • OpenAI reportedly plans to buy Google’s TPUs; Oracle will rent them out (Oct 27): The Information, via Reuters, reports OpenAI plans to buy Tensor Processing Units outright — a step beyond renting them through Google Cloud since 2025 — with no price, volume, or timing disclosed and neither company confirming. Separately, Oracle announced it will offer Google TPUs on Oracle Cloud Infrastructure from 2027, the first hyperscaler outside Google to do so. (Silicon Report, Skimwire)
  • Liquid AI’s LFM3: small models, six ways (Oct 27): a family of six compact models sharing one architecture across text, images, audio, and decision tasks — the week’s reminder that not every release is a frontier bid. (PYMNTS)