Vol.55 · November 2, 2026
dera news AI Weekly Vol.55 | 2026-11-02 - This Week's AI News
🤖 dera news AI Weekly Vol.55
2026-11-02
Anthropic made public its stock market listing filing (the prospectus) on October 26 (US time) and showed that it was profitable at the operating level on $11.6 billion in revenue for April to June. Alphabet said in its earnings on the 28th (US time) that its capital spending hit a record $51.4 billion, and its share price the next day had its biggest fall since its 2004 listing.
📊 What You Need to Know This Week
Alphabet's capital spending was so heavy that its free cash flow turned negative, and Anthropic plans to rent Broadcom chips from a dedicated company that buys them (stories 1 and 2). When you compare AI companies' numbers, look at whether each pays for its equipment itself or rents it.
OpenAI priced its mid-tier GPT-6.1 Terra at half of its top-tier Sol, and Google will double the price of its top Gemini 4 Carbon after an introductory period (stories 3 and 9). Google has not said when Carbon's introductory period ends.
Anthropic reported revenue that counts sales through cloud providers in full, and OpenAI asked outside experts to verify a proof its own AI produced (stories 1 and 4). Both are the companies' own announcements, so when you use them in internal documents, add how the figures were counted and the result of the verification.
CrowdStrike reported an attack in which AI agents (AI that carries out work on its own toward a goal) running mainly on DeepSeek models split up the work after intrusion, and Google gave its top Gemini 4 Carbon first to cyber defense organizations (stories 3 and 8).
The government has begun discussing whether to add penalties to the AI Promotion Act, and OpenAI, Google and Microsoft said they will read a common signal for refusing AI training (stories 5 and 7). Both are still to be decided: Japan's direction by the end of the year, OpenAI's support within a few months and Google's in 2027.
💡 This Week's Actions
1. Give GPT-6.1 Terra (story 9) one routine PC task and time it (30 min) If you are on one of the ChatGPT plans that include it (story 9), hand Terra one weekly task, such as moving the numbers in a PDF table into a spreadsheet. Record how many minutes it took compared with doing it by hand, and where it got stuck. → OpenAI: Introducing GPT-6.1 Terra (English)
2. Decide whether your company's site may be used for AI training, using the common robots.txt signal (story 5) (20 min) Decide with your PR or legal contact whether the articles and documents on your company's site may be used to train AI. Once decided, ask the site's administrator or the agency that built it to write it into robots.txt (the file that tells crawlers, the programs that visit sites automatically, which parts they may read) the way story 5 describes. The line has no effect until each company starts reading it. → ITmedia NEWS: Refusing AI training in one line of robots.txt (Japanese)
3. Prepare for a price rise like Gemini 4 Carbon's (story 3) by setting the most you will pay for each AI service (30 min) List the AI services you pay for now, and write down how much of a price rise you would accept and what you would cut if prices went up. Share the list with whoever owns the budget. → Google: Introducing Gemini 4 Carbon (English)
📰 This Week's AI Articles (10 stories)
1️⃣ Anthropic makes its listing filing public, with an operating profit for April to June
🏷️ Funding, IPO, Anthropic What happened? On October 26 (US time) Anthropic made public its listing filing (S-1, the prospectus) with the US Securities and Exchange Commission (SEC) and said it will list on Nasdaq. Revenue for April to June was $11.6 billion, and operating profit, which shows the earnings from the core business, was about $974 million. The revenue counts sales through cloud providers such as AWS in full. After listing, the seven co-founders will still hold one special share (Class F) through a joint company, and with it 50.1% of the voting rights. The Wall Street Journal reports, citing people familiar with the matter, that Anthropic is aiming to list in the week of November 16 and could raise more than $100 billion at a valuation of $1.5 trillion to $2 trillion (the timing may change). Our take The same week, three banks and the US investment firm Blackstone announced they will arrange $63 billion for a dedicated company that buys chips from the semiconductor maker Broadcom and rents them to Anthropic. According to the Financial Times, Blackstone expects Anthropic's credit rating after listing to be in the "BB" range, just below investment grade. If you plan to build Claude into next year's budget, assign someone to check Anthropic's earnings and credit rating once a year, as you would for a major supplier. 📎 Read the original
2️⃣ Alphabet's share price has its biggest fall since listing, as investors question AI capital spending
🏷️ Earnings, Funding, Google, Japan What happened? Alphabet announced its July-to-September earnings on October 28 (US time). Google Cloud's revenue was $30.7 billion, more than double the year before. Capital spending hit a record $51.4 billion, and free cash flow (the cash earned from operations minus capital spending) was negative $9.1 billion. In trading on the 29th the share price fell about 12.6%, its biggest drop since the 2004 listing. In Tokyo the Nikkei average fell as much as 3% on the 30th, and Advantest, which makes chip testing equipment, fell as much as 9.1% even though it had raised its forecast on the 28th. Our take Last week Amazon's Andy Jassy said AWS capacity for 2027 is almost sold out, but this week Alphabet's shares were sold and investors looked hard at when AI spending will be recovered. AI companies are in a hurry to recover it, so today's discounts and free tiers are unlikely to last on the same terms. When you sign an annual contract for an AI service, check that the contract says how long the price is fixed. 📎 Read the original
3️⃣ Google's top "Gemini 4 Carbon" goes first to cyber defense organizations
🏷️ AI Models, Security, Google, Pricing What happened? On October 28 (US time) Google announced its top model, "Gemini 4 Carbon," and began offering it first to selected organizations that handle cyber defense. A month earlier it had also given Gemini 4 Argon to cyber defense organizations first. Google says it will open Carbon to developers and the general public "as soon as possible," after gathering feedback and improving the safeguards, but gave no date. It starts at $2 per million tokens (the unit AI uses to count the text it reads and writes) for input and $10 for output, and after an introductory period whose end date it has not given, the prices rise to $4 and $20. On an overall index the evaluator Artificial Analysis revised on the 27th, it took first place, narrowly ahead of Anthropic's Opus 5.5. Our take There is no date for when ordinary users can use it, and for the top models, whether you can use one is likely to depend more and more on whether the provider selects you. Ask the vendors of your security products and services whether they take part in early-access programs like this. 📎 Read the original
4️⃣ OpenAI says its AI overturned the "Exponential Time Hypothesis" in computer science
🏷️ Research, OpenAI What happened? On October 26 (US time) OpenAI announced that its internal AI model had overturned the "Exponential Time Hypothesis" in computer science. The hypothesis holds that certain hard problems cannot be solved much faster than by trying every combination, and it has been widely believed since the late 1990s. OpenAI published a paper and a proof written in Lean, software that checks a proof one step at a time by machine. It asked outside experts to verify it independently, including Russell Impagliazzo and Ramamohan Paturi of the University of California San Diego, who proposed the hypothesis, and Ryan Williams of MIT. Williams says he is still checking whether the proof holds. Our take Three weeks ago OpenAI published a batch of AI-written math papers and later withdrew some of them. This time it asked outside experts to verify at the same time as the announcement, and whether the result is correct will not be known until that is done. Before you use a conclusion an AI researched in an internal decision, open one of the source documents and check that it says what the AI says it does. 📎 Read the original
5️⃣ OpenAI and two other companies say they will honor a common robots.txt signal for refusing AI training
🏷️ Data, Copyright, OpenAI, Google, Microsoft, Anthropic What happened? On October 30 (US time) OpenAI, Google and Microsoft said they will support a common robots.txt signal that lets a site refuse use of its content in AI training all at once. The signal is a draft that Kevin Kelly, who works on standards at Anthropic, submitted in a personal capacity to the IETF (the body that sets internet technical standards), and it is not yet a standard. In the User-Agent line, which names who the rules are for, you write "__usage_ai-training," then Disallow to refuse training or Allow to permit it. OpenAI will support it within a few months, Google in 2027, and Microsoft's crawler will read it from February 1, 2027. Our take Until now, site operators have had to look up the crawler name of each AI company and write a refusal for each. Once the companies start reading the signal, one line can state their intent on training. Until then the signal alone does not refuse training, so keep the per-company refusal lines you have now. 📎 Read the original
6️⃣ Microsoft offers voluntary exit to about 60,000 employees, about a quarter of its workforce
🏷️ Employment, Microsoft What happened? On October 26 (US time) Microsoft offered voluntary exit packages to about 60,000 employees in the US, UK, Canada and India. Eligible employees are those below vice president level in the participating divisions, about 27% of its roughly 223,000 employees. Employees decide by early December whether to accept, and a spokesperson says the terms are comparable to standard severance in each region. In an internal memo, Chief People Officer Amy Coleman said it gives employees who are thinking about their future outside the company a new choice, and gives participating divisions another way to adjust staffing to changing needs. The spokesperson says AI is not the reason for this offer. Our take Microsoft does not cite AI as the reason, but it has shown about a quarter of its employees the option to leave. Take time now to talk through which of your team's tasks will go to AI and which people will keep doing. 📎 Read the original
7️⃣ The government starts discussing a review of the AI Promotion Act, with penalties for malicious operators the focus
🏷️ Policy, Japan What happened? On October 26 the government held an "expert panel on AI policy" and began discussing a review of the AI Promotion Act, which was enacted in 2025. The focus is whether to add penalties for malicious operators. The act has no penalties now, and Digital Minister Toshiharu Furukawa asked for views on "how trust should be ensured." The government lists the risk of AI "running out of control" and acting against human intent, and the growing seriousness of cyberattacks that abuse AI, among the matters to consider. A working group of university professors and others will sort out the issues from the 28th, hear from AI developers and set out the direction of the review within the year. Our take A month ago, on October 2, Furukawa said he would consider changes including an amendment, and now the discussion has begun. It is not yet clear whether penalties would apply only to companies that build AI or also to companies that use AI to offer services, and how that is written will change the impact on the many companies in Japan that use AI a great deal. If you offer services that have AI built in, ask your legal contact to watch how the scope of "operators" is written in the direction due within the year. 📎 Read the original
8️⃣ A China-linked attack group targets more than 200 organizations with its AI tool "ARTEX"
🏷️ Security, China, DeepSeek, Agents What happened? On October 29 (US time) the security firm CrowdStrike reported that the China-linked attack group "VAPOR PANDA" used its AI tool "ARTEX" to target more than 200 organizations over the 11 days from October 17. ARTEX ran mainly on models from the Chinese AI company DeepSeek, several AI agents split up the investigation after intrusion and the moves to other servers, and a human gave instructions at key points. Post-intrusion activity was confirmed at nine organizations, and the targets include foreign ministries in various countries and US semiconductor and defense organizations. CrowdStrike says this attack is separate from the ARTEX attack Microsoft reported on October 21, and does not say whether Japanese organizations were among the targets. Our take Last week GMO Cybersecurity by Ierae found ARTEX at the base of attacks aimed at Japanese companies, and this time CrowdStrike reported an attack that ran the same tool on DeepSeek models. CrowdStrike has also confirmed that financially motivated attackers and others are using ARTEX or similar AI setups. If information may have leaked, CrowdStrike recommends replacing signing keys and service credentials, including ones left in old backups, so list now what you would need to replace. 📎 Read the original
9️⃣ OpenAI's PC-operating "GPT-6.1 Terra" costs half of the top model
🏷️ AI Models, Agents, OpenAI, Pricing What happened? On October 26 (US time) OpenAI released the mid-tier model "GPT-6.1 Terra." It is strong at "computer use," where the AI looks at the screen and moves the mouse and keyboard. Users on ChatGPT's Plus, Pro, Enterprise and Edu plans can use it in ChatGPT Work and Codex (OpenAI's coding tool), and the API (the doorway for calling the AI from software) costs $1 per million tokens for input and $5 for output, half of GPT-6.1 Sol. In OpenAI's own evaluation, Terra finishes computer-use tasks more than three times as fast as the higher-tier Sol and Astra, but OpenAI recommends Sol or Astra for the most complex coding. Our take Last week The Information reported that OpenAI gave discounts of over 30% on contracts worth more than $100 million a year, and this week it released a model cheaper than the previous generation's GPT-5.6 Terra. If, as reported, OpenAI is competing for customers on price, the cost of handing PC work to AI is likely to keep falling for a while. Decide with your IT contact in advance which apps and folders the AI may operate. 📎 Read the original
🔟 Apple's Siri AI comes to Japanese, on iPhone 15 Pro and later
🏷️ Apple, Japan What happened? In iOS 27.2, released on October 26 (US time), Apple extended its Siri AI features, "Siri AI," to Japanese, French, Korean, Portuguese and Spanish. It works on iPhone 15 Pro and later, and remembers past conversations so you can ask again about what it answered before. It is not offered in the EU or mainland China. Our take Apple released Japanese Siri AI in October, as it announced in September. If you have an iPhone 15 Pro or later, ask the Japanese Siri AI to summarize one long email and see whether it is usable for work. 📎 Read the original
📚 Editor's Note
In the risk factors section of its prospectus (the list of dangers shown to investors), Anthropic wrote that it recently delayed the start of training its latest large model by more than a month. It explains that this was to strengthen protections against theft of the model's contents (its weights) and its intellectual property. That section is where you can check how Anthropic itself sees the reasons Claude's service or new model releases could stop or be delayed.
See you next week, with useful information and something to think about. The dera news team