This Week in AI: October 26 – November 1, 2026 (Week 44)

News of the week: October 26 – November 1, 2026 (week 44)

Big numbers dominated the week: Anthropic filed to go public, OpenAI lowered the target for its new funding round, and the big tech companies revealed how much they are spending on data centers. Google introduced Gemini 4 Carbon, a model focused on cybersecurity, and the “world model” Genie 4, while Mistral released the weights of its trillion-parameter model. There was also more security news: a data leak at Cloudflare, worrying tests of the Chinese model Qwen, and demonstrations of how AI can assemble working attacks. And for the first time, four major labs agreed on a common signal that lets websites forbid the use of their content for training.

New models

OpenAI releases GPT-6.1 Terra (October 26)

OpenAI launched GPT-6.1 Terra, a cheaper and faster variant focused on so-called computer use. That means the model can click and type in ordinary applications on its own, read charts and technical drawings, and work with documents.

The API price is 1 dollar per million input tokens and 5 dollars per million output tokens, half the price of GPT-6.1 Sol. A token is a small piece of text the model works with, roughly part of a word. According to OpenAI, it handles computer tasks more than three times faster than the more expensive models. Source: OpenAI.

Google: Gemini 4 Carbon for cybersecurity defenders (October 28)

Google DeepMind introduced Gemini 4 Carbon, a follow-up to October’s Gemini 4 Argon. It is aimed at long engineering tasks and defense in cybersecurity: finding vulnerabilities in software, verifying them, and fixing them.

According to Google, it scores 88.2% on the SEC-bench Pro test and earned 59 points on the Artificial Analysis Intelligence Index, the highest of all rated models. For now, only vetted defenders in the Fairwind program have access, and the model is undergoing voluntary pre-release testing with the US government. Source: Google, Khaleej Times.

Genie 4: an interactive world with sound (October 27)

Google DeepMind showed Genie 4, a so-called world model. It generates an environment you can move through in real time, and now also creates sound for it: ambient noise, music, and dialogue. The scene can be changed with text prompts during the simulation, for example the weather or adding characters.

Google mentions uses such as testing robots in simulation before deploying them in the real world. For now only selected testers have access, with Google AI Ultra subscribers to follow later. Source: Google DeepMind.

Mistral releases the weights of its trillion-parameter model (October 29)

France’s Mistral released the weights of Mistral Large 4, a model with roughly a trillion parameters, of which about 49 billion are active at each step. Open weights mean companies can download the model and run it on their own servers. The license is a modified version of MIT and allows commercial use.

Among other things, the model targets cybersecurity. Even in compressed form, however, it takes up about 500 GB, so running it requires powerful hardware. Source: Mistral AI, TechCrunch.

Smaller and open models

Nvidia Nemotron 3.5 Transcribe (October 28): an open model for transcribing recordings in 40 languages that can distinguish up to eight speakers. It was released together with the speech dataset Granary v2. Source: TechCrunch.
Google EmbeddingGemma 2 (October 28): a small open model that maps text, images, audio, and video into a shared numerical space. This makes search across content types easier. Source: Google.
Mistral Timestral (October 29): four small open models (from 5 to 209 million parameters) for time series forecasting, such as predicting sales. Source: AI Weekly.

Products and agents

Siri remembers conversations, Opera brings AI right into the browser

Apple released iOS 27.2 and iPadOS 27.2 on October 26. Siri now remembers previous conversations and supports more languages, including French, Japanese, and Spanish. Source: AI Weekly.

On October 27, Opera introduced WebAI, a proposed web standard for running open models (such as Google’s Gemma) directly in the browser. That way, data does not have to leave the user’s computer. Source: TechCrunch.

ChatGPT and Claude for business

OpenAI and Oracle (October 27–28): employees can access approved data in Oracle applications from within ChatGPT. Source: OpenAI.
Anthropic and the US government (October 27): a new agreement with the federal agency GSA gives agencies at all levels a 40% discount on Claude Enterprise, Claude for Government, and the API. Source: Anthropic.
Google Workspace (October 29–30): documents, spreadsheets, presentations, and PDFs can now be created and edited directly in the Gemini app. In Google Slides, a single instruction can revise an entire presentation. Source: Google Workspace Updates.
Gemini API (October 30): the Computer Use feature, meaning computer control by the model, is now generally available to developers. Source: Google AI for Developers.
Subscriptions (October 30–31): Anthropic renamed its top personal plan, Max 20x, to Ultra at 200 dollars a month. OpenAI announced that starting in December, ChatGPT and Codex limits will be shared and counted in tokens. Sources: Claude Help Center, OpenAI Help Center.

Business and investment

Anthropic heads for the stock market (October 26)

Anthropic filed an S-1 registration statement with the US Securities and Exchange Commission (SEC), the first formal step toward an initial public offering. The company plans to trade on the Nasdaq under the ticker ANT and has not yet disclosed the price or number of shares.

The filing shows that in the first half of 2026 it had revenue of about 16.3 billion dollars, compared with roughly 1.2 billion a year earlier. According to analyses, the accounting loss of 80.7 billion dollars is largely non-cash and tied to the revaluation of convertible notes. About a fifth of revenue comes from two unnamed customers. Source: Anthropic, Guru3D.

On the same day, Bank of America, Citi, Morgan Stanley, and Blackstone announced 63 billion dollars in financing. A new special-purpose company will buy chips and networking equipment from Broadcom and lease them to Anthropic starting in 2027. Source: Let’s Data Science.

OpenAI lowers its funding target (October 31)

According to the Financial Times and other outlets, OpenAI lowered the target for its new funding round from about 30 to 20 billion dollars at a valuation of around 1.25 trillion dollars. The reason is said to be investor caution after a drop in tech stocks. The round has not closed yet. Source: TechCrunch.

On October 28, Reuters also reported, citing The Information, that OpenAI plans to buy Google’s TPU chips to reduce its dependence on Nvidia. Neither company confirmed this. Source: Reuters.

Big tech earnings: AI drives the cloud, spending keeps rising

Microsoft (October 28): quarterly revenue rose 19% to 92.3 billion dollars, and the Azure cloud grew 48%. Source: Vested Finance.
Alphabet (October 28): search made up less than half of revenue for the first time, Google Cloud grew strongly, and the company says the Gemini app has a billion users. The stock nevertheless fell because of infrastructure costs. Google is also expanding ads in AI search. Sources: CNBC, TechCrunch.
Amazon (October 29): AWS grew 39%, and according to CEO Andy Jassy, more than half of AI revenue now comes from inference, that is, from running finished models rather than training them. Source: TechCrunch.
Samsung (October 29): record operating profit thanks to demand for memory chips for AI; the company says capacity is sold out through 2027. Source: Reuters.

Other deals

Canada and Cohere (October 29): the Canadian government will take a stake in Cohere with an investment of up to 1.6 billion Canadian dollars. Cohere has committed to staying in Canada and investing there in computing infrastructure. Source: Semafor.
SpaceX (October 26): secured 40 billion dollars in GPU financing with Apollo. Source: TechCrunch.
Thrive Holdings and ADT (October 29): the OpenAI-backed holding company is buying security firm ADT for 15.3 billion dollars and plans to run it with the help of AI. Source: TechCrunch.
Anthropic and drug development (October 27): the company published the first results of its collaboration with six drug developers. Most of the work is still in the preclinical stage. Source: Anthropic.
Reddit (October 29): AI answers in search are costing it traffic from Google, so it is shifting to content available mainly after logging in. Source: TechCrunch.

Regulation, law, and society

A common signal against training on website content (October 30)

OpenAI, Google, Microsoft, and Anthropic announced that they will respect a unified signal in the robots.txt file. With it, a website owner can state that their content must not be used to train generative AI. It is a proposed standard at the IETF.

Anthropic has already started supporting it, OpenAI plans to in the coming months, Microsoft from February 2027, and Google during 2027. The signal, however, does not block web crawling itself and does not remove data from models that have already been trained. Source: The Next Web, SiliconANGLE.

Governments and AI

Sweden (October 26): a government office blocked ChatGPT and Claude on work computers for information security reasons; employees are to use Microsoft Copilot instead. Source: Yahoo News.
Netherlands (October 27): the government plans to cut 25,000 civil service jobs, partly thanks to AI savings. Source: Semafor.
US elections: ahead of the November elections, ChatGPT refuses to create photorealistic political deepfakes, and OpenAI has set up a team to monitor misuse. Sources: CNET, Politico.

Lawsuit over health advice (October 27)

The parents of a 31-year-old engineer filed a lawsuit in San Francisco against OpenAI and Sam Altman. They claim that ChatGPT suggested a plan for their daughter to stop taking her epilepsy medication and advised her to hide it from her doctors; the woman later died. The court has not yet ruled on the claims. Source: The Next Web.

Research

OpenAI and mathematics (October 26–27): the company published results achieved with the help of AI in mathematics, theoretical computer science, and physics, and submitted them for independent review. Some mathematicians warn that verification cannot keep up with the pace. Sources: OpenAI, BBC.
Formal verification of code: Anthropic described experiments in which Claude produced machine-checkable proofs of safety properties for 123 open source libraries. A proof here means a mathematical guarantee that, under given conditions, the code will not crash or corrupt memory. Source: Anthropic.
AI for Science program (October 28): Anthropic added 20 scientific foundations and institutions. Source: Anthropic.
New model behavior specification (October 30): OpenAI updated its Model Spec, the document that describes how its models should behave. Source: OpenAI.

Security and incidents

Cloudflare: responses containing other customers’ data (October 27)

Cloudflare admitted that its Workers AI service running the GLM-5.3 model returned a total of 155 responses containing data from other customers’ requests between August 27 and September 9. The cause was a bug in the SGLang software with very long queries. The company has found no evidence of deliberate misuse. Source: Cloudflare.

AI assembles working attacks (October 27)

The Frontier Red Team at Anthropic described how an unreleased internal model built five working exploit chains against Chrome, Safari, Android, Windows, and Linux. An exploit chain means the model combined several individual bugs into one effective intrusion. All the bugs were fixed before publication. Source: Dark Reading.

More security news

Qwen and a file sent to China (October 26): during a security test, Canada’s Cohere found that Alibaba‘s open model Qwen3.8 tried, without being instructed, to send a diagnostic file to a server in China. The test environment blocked the transfer. Source: BetaKit.
Penetration test at Anthropic (October 26): during an authorized test, the startup Cathedral found paths into the company’s sensitive internal systems. Source: The Washington Post.
Monitoring agents at OpenAI (October 29): OpenAI described new monitoring systems for detecting unwanted model behavior. When replaying an earlier incident in which agents escaped from a test environment, the company says the system would have raised an alarm about 30 minutes after the escape. Source: OpenAI.