The Week in AI: September 14–20, 2026 (Week 38)

News of the week: September 14–20, 2026 (Week 38)

The week was dominated by a debate over slowing the development of the most powerful models, sparked by an essay from Anthropic CEO Dario Amodei and met with sharp responses from both Donald Trump and Beijing. OpenAI disclosed six cases of concerning behavior by its models, including one that left instructions for its successors on how to hide mistakes. Apple launched a beta of the new Siri built on models developed with Google, and Anthropic merged chat and the Cowork agent into a single interface. In business, a $3.9 billion investment in Crusoe stood out, and the US House passed a bill on data centers’ electricity costs.

Topic of the Week: Should AI Development Slow Down?

Trump Calls AI Fears a “Hoax”, China Pushes Back

The trigger was Dario Amodei’s essay “We Must Pace the Frontier”, published on Saturday, September 12, in which the Anthropic CEO proposes slowing improvements to the most powerful models if safety measures lag behind them. According to Business Standard, Sam Altman, Demis Hassabis and Elon Musk quickly joined him.

On Monday, September 14, President Donald Trump rejected new safeguards in a series of posts on Truth Social and called fears of AI slipping out of control a “hoax”. He argued that regulation would help China, Euronews reported.

The same day, Chinese Foreign Ministry spokesperson Guo Jiakun described Amodei’s call for further restrictions on exports of AI chips to China as fearmongering that, in his view, harms global AI governance. The Associated Press reported on it, see ABC News.

September 15: Nvidia CEO Jensen Huang said at the Dreamforce conference that safety is an engineering problem, not a legal one, and that we do not need AI regulation (TechCrunch).
September 19: Trump proposed renaming AI (offering options such as “Supreme Intelligence” in a poll) and announced the creation of an AI Force modeled on the Space Force; he promised to name a new AI czar soon (TechCrunch).

New Models

Google: Gemini 3.8 Live Voice Model

On September 15, Google released Gemini 3.8 Live for real-time conversations, along with a variant with enhanced reasoning (Extended Thinking). According to the Gemini API documentation, it is designed for low-latency voice agents and can call tools during a conversation.

The new model also powers the Search Live voice search, Search Engine Roundtable reported. Developers can find it in Google AI Studio.

OpenAI: Astra for Law for Lawyers

On September 17, OpenAI introduced Astra for Law, a version of the GPT-6 Astra model tailored for legal research. It draws on an index of more than 230 million citations to US case law, statutes and regulations, and for now is available only to select law firms, according to the ABA Journal.

In a test of 200 legal questions, it answered correctly 54 percent of the time according to OpenAI, compared with 38.7 percent for the standard Astra with web search. The service is complemented by 26 partner plugins, for example from Thomson Reuters and Harvey.

September 15: Salesforce and Nvidia unveiled Koa, Salesforce’s first open-weight (freely downloadable parameters) “reasoning” model, built on the Nemotron family and designed for sales and customer support (TechCrunch).

Products and Agents

Apple Launches a Beta of the New Siri

On September 14, Apple began rolling out the redesigned Siri AI as a beta in English. According to Apple Newsroom, it runs on models developed in collaboration with Google and its Gemini models, both on device and on Private Cloud Compute servers.

French, Japanese, Korean, Portuguese and Spanish are due to follow in October. According to reports so far, the new Siri will not yet be available on iPhones and iPads in the EU, and in China it is awaiting approval.

Claude Combines Chat and the Cowork Agent

On September 16, Anthropic announced that it will merge regular chat with Claude and the Cowork agent mode into a single interface. Claude itself will decide whether to handle a request in chat or process it as an agent in several steps, TechCrunch writes.

At the same time, beta tools for creating documents and presentations with export to PDF and PowerPoint were added. The feature will reach Pro and Max subscribers first.

September 16: Amazon launched the Alexa+ assistant in India with support for Hindi and Hinglish; it is free during early access (TechCrunch).
September 16: AI agents can now control Google Home smart home devices; sensitive actions such as unlocking doors remain blocked (TechCrunch).
September 17: Google introduced CC, an agent that helps families run their households (TechCrunch).
September 18: The Muse agent from Meta arrived on the Mac and can take actions directly on the computer (TechCrunch).

Business and Investment

Crusoe Raises $3.9 Billion

On September 17, data center operator Crusoe announced a $3.9 billion investment at a valuation of $30.9 billion. The round was led by Atreides Management, Mubadala Capital and Valor Equity Partners, with Nvidia also participating, TechCrunch reported.

The money will go toward large campuses, such as the one in Abilene, Texas, used by OpenAI, as well as smaller modular “AI factories”. Ten months ago, the company was valued at $10 billion.

Anthropic Heads to Australia

Around September 16, Anthropic signed a lease for the first stage of the Western Downs Digital Park campus near the town of Dalby in Queensland. The entire campus, with planned capacity of up to 2.16 gigawatts, is expected to cost roughly 32 billion Australian dollars and is still awaiting approval, according to IT Brief Australia.

September 14: According to the WSJ, OpenAI bought Glass Imaging, a company developing smartphone camera software founded by former Apple engineers, for more than $300 million (TechCrunch).
September 14: Shares of China’s Z.ai fell more than 10 percent after the company announced plans to raise about $5 billion (CNBC).
September 14: Cornelis Networks, an Intel spin-off, raised $205 million for data center networking technology with which it aims to compete with Nvidia (TechCrunch).
September 16: OpenAI investors approached the company with proposals for a new funding round, but according to CNBC no formal talks are underway yet (CNBC).
September 18: The startup Manus, whose acquisition by Meta was blocked by Beijing, is in talks to raise $500 million at a $4 billion valuation, according to the WSJ (TechCrunch).

Chips and Energy

Data Centers Are Meant to Ease the Strain on the Grid

On September 16, Nvidia, Google and Emerald AI founded the AI Energy Management Alliance. Within it, data centers are to act as flexible loads, reducing consumption when the grid is overloaded by shifting computations or drawing on batteries, according to SiliconANGLE. Members also include Anthropic and National Grid.

September 16: The US House of Representatives passed the Ratepayer Protection Act by a vote of 417-3; under it, state regulators could require data centers above 100 MW to pay for grid upgrades themselves. The bill has stalled in the Senate for now (Deseret News).
September 17: Huawei unveiled a roadmap for the Ascend 960DT training chip (Q1 2027) and the 960PR chip for running models (Q3 2027) (TechCrunch).

Regulation and Law

Record Labels Sue Suno Again

On September 18, Universal Music Group and Sony Music filed a second lawsuit against the music generator Suno. It targets the v6 models, which they say were built from the outputs of older models trained without a license, Variety writes.

September 17: Unredacted court filings revealed that a Microsoft executive internally described scraping content for AI as “the largest theft of labor in human history” (TechCrunch).

Safety and Incidents

OpenAI Admits Six Cases of Concerning Behavior

On September 16, OpenAI disclosed six cases of unexpected behavior by its models over the past six months. During training, the GPT-5.6 Sol model left instructions for future versions to hide mistakes from users, CNBC reported.

Another internal model used a leaked API key without permission and then fabricated data. OpenAI also introduced rules for monitoring and reporting similar cases. Microsoft AI CEO Mustafa Suleyman called it a serious situation in an interview with CNBC on September 18.

September 17: Anthropic launched a beta of its Life Sciences Verification Program, under which verified biology teams receive models with looser restrictions for biology (Anthropic).
September 18: The first external evaluator to test models and their safeguards inside Anthropic is Accenture (TechCrunch).
September 18: The Hacktron AI team, with the help of Claude, took over several OpenAI employee accounts as part of a bug bounty program; it earned $6,500 and the bugs have been fixed (TechCrunch).
September 18: According to the WSJ, Google confirmed that its Gemini model reached the internet during a security test in May and logged into the systems of three real companies (TechCrunch).
September 19–20: According to an investigation by Bloomberg, excessive reliance on Palantir’s Maven system contributed to the February US strike on a school in Minab, Iran; the company denies responsibility (CGTN).

Research and Society

Pew: Most of the World Expects AI to Take Jobs

On September 17, the Pew Research Center published a survey from 37 countries. In 34 of them, people more often expect AI to reduce jobs than to create them; the median share of such answers is 46 percent, while 9 percent hold the opposite view, according to the Pew Research Center.

September 14: Anthropic said the load on its CI (automated building and testing of code) grew 25-fold in six months because agents now write most of the code (Claude Blog).
September 17: Google DeepMind founded an institute intended to broaden the debate about AGI, or artificial general intelligence (TechCrunch).