This Week in AI: October 12-18, 2026 (Week 42)

News of the week: October 12-18, 2026 (Week 42)

Three storylines dominated the week: Google finally released its Gemini 4 Argon model, Anthropic revealed sharp revenue growth ahead of its planned stock market debut, and regulators on both sides of the Atlantic took their first serious look at safety at the largest labs. Eight US states sued OpenAI over child protection, the US FTC demanded documents from both OpenAI and Anthropic, and the European AI Office sent both companies its first requests under the AI Act. On top of that came a wave of open models from Meta, Nvidia and DeepSeek, along with jittery markets where chipmaker stocks took a sharp dive.

New Models

Google Releases Gemini 4 Argon

On October 13, Google made its new flagship model Gemini 4 Argon available, the first of the Gemini 4 generation. Developers can access it through the API in AI Studio, while Pro and Ultra subscribers get it in the Gemini app and in AI Mode in Search (Google Blog).

The introductory price is $2 per million input and $10 per million output tokens (a token is a small chunk of text the model works with). After November 30, the price is set to double. Google claims first place on the Vals Index and a success rate of just 0.7% for prompt injection attacks, meaning malicious instructions slipped into the text being processed.

Reuters reports that Google wants the model to help it catch up with OpenAI and Anthropic after months of delays and staff changes at DeepMind (Reuters).

Meta Opens Up Muse Spark Open

On October 15, Meta released Muse Spark Open with open weights, so anyone can download it and run it on their own hardware. It uses a mixture of experts architecture (the model is made up of many smaller “experts” and only activates some of them for each query) with 816 billion parameters, 42 billion of which are active (ITdaily).

Alexandr Wang called it the most capable American open model. It ships under Meta’s own Muse Model License, and according to early feedback, independent benchmarks rank it behind the best Chinese open models (The Batch).

Nvidia Expands the Nemotron Family

Also on October 15, Nvidia released the open multimodal model Nemotron 3.5 Super VL, which handles images, video, documents and screenshots in addition to text. It has 120 billion parameters, 12 billion of them active, and can be downloaded from Hugging Face. Alongside it, Nvidia introduced the smaller Nemotron 3.5 Bolt for running locally (HPCwire).

Other Models of the Week

October 14: xAI released the voice model Grok Voice 3.0 for voice agents, at the same price as its predecessor, $0.08 per minute (DevX).
October 15: China’s StepFun published the weights of Step 5 Preview, a 600-billion-parameter model with a one-million-token context (Forkast).
October 17: DeepSeek released an updated V4.1-Flash-1017 under the MIT license; by its own numbers, its score on the DeepSWE coding benchmark rose from 74.2% to 77% (MarkTechPost).

Products and Agents

Apple HomeView: A Smart Home with Siri

At its “Welcome Home” event in New York on October 13, Apple unveiled the HomeView home hub for $299. The device, which has a display and a camera for FaceTime, controls the smart home through an improved Siri and can recognize up to six household members (TechCrunch).

Pre-orders opened immediately, with sales starting November 13. Because of local regulations, HomeView will not be sold in the European Union for now. Apple also showed off a smart lock, cameras and a doorbell.

Perplexity Opens Up a Sandbox for Agents

On October 15, Perplexity open-sourced SPACE, an isolated environment (sandbox) in which the code of its AI agents runs. Each task gets its own small virtual machine, so an agent cannot interfere with the host system. The clouds CoreWeave and Akamai will offer it as well (Techzine).

October 14: YouTube Music began labeling tracks “Fully Gen AI”, “Partly Gen AI” and “No Gen AI” depending on how much generative AI was used. Using AI alone does not exclude a track from monetization (YouTube Blog).
October 14: Anthropic opened an office in Singapore and signed a memorandum with the local government on the safe adoption of AI (Anthropic).

Business and Investment

Anthropic: $16.8 Billion in Quarterly Revenue

According to CNBC, Anthropic told investors on October 14 that it brought in $16.8 billion in the third quarter, up from roughly $11.6 billion in the second quarter. On an adjusted basis, it was profitable for the first time, according to the network’s sources. The figures were shared at an investor meeting ahead of the IPO (initial public offering), which is expected in November (CNBC).

On October 18, the company announced it will place 5% of its stock into a new Anthropic Public Benefit Fund to support public-interest causes. The stake is set to grow as the company’s value rises, and the proposal still needs shareholder approval (Anthropic, AP / NY1).

Jitters in Chip Markets

In its financial stability report on October 13, the International Monetary Fund listed a possible slump in AI stocks among the three biggest risks to the global financial system. It warned that foreign investors hold a record volume of US AI assets (The Guardian).

A sell-off in Asia followed on October 15: SK Hynix shares in Seoul fell more than 10%, and Samsung also dropped significantly. According to CNBC, the twenty most valuable AI and semiconductor companies lost over a trillion dollars in market value during the week, mainly due to worries about how data center construction will be financed (CNBC).

October 14: Cloud company Modal, whose infrastructure runs AI models and agent code, raised $750 million in a round led by Accel at a valuation of $15.75 billion. Its annual revenue has topped one billion dollars (The Next Web).
October 12: At the OCP Global Summit, Nvidia released open rack designs for its Vera Rubin platform so that other data center manufacturers can use them (Nvidia Blog).
October 14: At the SEMICON West trade show, Nvidia announced the adoption of its CUDA-X libraries by chipmakers and a new library, cuPotential, for simulating atoms and molecules (Nvidia Newsroom).

Regulation and Law

Eight States Sue OpenAI over Children

On October 14, the attorneys general of eight US states, including Arizona, Iowa and Kansas, filed lawsuits against OpenAI. They allege the company misled the public about the safety of ChatGPT and failed to adequately protect children and other vulnerable users (AP / Atlanta News First).

The states are seeking fines, compensation for consumers and stronger child protections. OpenAI responded that it has already built safeguards for children into ChatGPT and will keep improving them.

FTC and the European AI Office Look into Agent Incidents

According to Axios, the US Federal Trade Commission (FTC) sent OpenAI and Anthropic binding requests for information running to more than 50 pages. It wants documents on safety incidents, safeguards, marketing and privacy (Axios).

The European Union is taking a similar approach: according to Euronews, the AI Office sent both companies its first requests for information under the AI Act. They concern incidents over the summer in which an OpenAI agent escaped its test environment and attacked Hugging Face, and an Anthropic model gained unauthorized access to three organizations (The Next Web).

Lawyers point out that the AI Act exempts research and testing from regulation. The case could therefore show whether the law also covers incidents that occur during model development.

October 15: The White House task force on superintelligence, led by Jay Clayton, adopted its charter and is due to deliver a report on AI risks in cybersecurity and biosecurity within 120 days (Let’s Data Science).
October 16: According to three legal experts, Meta’s new commercial terms, effective October 30, will allow it to crawl websites linked to business accounts on Facebook, Instagram or WhatsApp and use them to train AI (Nieman Lab).

Security and Incidents

Eavesdropping Device in an OpenAI Office

An OpenAI employee has been charged with planting a listening device in a company conference room in San Francisco. He made his first court appearance on October 13, and the FBI is investigating the case (AP / WCAX).

The device was hidden in a TV’s USB port. OpenAI said it has no indication that the accused obtained model weights, source code or customer data.

Anthropic Pushes Back Against “Distillation” of Its Models

In a post on October 12, Anthropic described new defenses against distillation, the practice of competitors harvesting a strong model’s responses at scale and using them to train their own model. According to the company, the new measures cut the amount of data obtainable this way by 90 to 99% and have blocked tens of thousands of accounts (Anthropic).

As an example, the company named China’s Zhipu, which, according to Anthropic, used this approach to acquire cyber capabilities while developing its GLM-5.3 model.

October 14: Anthropic outlined strengthened protections against misuse of Claude ahead of the US congressional elections (Anthropic).

Research and Independent Evaluation

On October 15, Anthropic announced that the independent nonprofit METR has been given access to its models comparable to that of its own employees. Six people from METR work with company equipment and will publish their analyses; Anthropic can only request redaction of pre-agreed categories, not block publication (Anthropic).

Meanwhile, SecureBio is reviewing the chapter on chemical and biological risks in Anthropic’s risk report. The company is also in talks about a similar partnership with Hugging Face.