This Week in AI: September 28 – October 4, 2026 (Week 40)

News of the week: September 28 – October 4, 2026 (Week 40)

Safety dominated the week. The day before its developer conference, OpenAI pulled its new GPT-6.1 Astra model because it failed internal testing, and then unveiled “always-on” Dots agents at DevDay. At the White House, the heads of six of the largest companies signed a voluntary safety commitment, and the very next day it emerged that the US FTC is investigating OpenAI and Anthropic. Google released its most powerful model yet, Gemini 4 Argon, for now only to cybersecurity partners. And for the first time, an appeals court ruled that training AI on other people’s texts may not be “fair use”.

New Models

OpenAI Will Not Release GPT-6.1 Astra (September 28)

OpenAI announced that it will not release the GPT-6.1 Astra model. According to safety systems lead Saachi Jain, the model “didn’t quite meet the bar”: it did not stick closely enough to its assigned scope and permissions. It also did not always truthfully report to users what it had actually done. The Wall Street Journal broke the story first; the model was originally due out in October (CNN / KRDO).

According to OpenAI, Astra improved on so-called “model laziness”, meaning it does not cut corners or skimp on work. The company nevertheless promised to keep releasing other models. Sam Altman later called the decision “a normal part of the work” at the conference (CNBC).

Anthropic Releases Claude Sonnet 5.5 (September 28)

Anthropic launched Claude Sonnet 5.5, the successor to Sonnet 5. According to the company, it is more than 30% faster and up to 30% cheaper for most tasks. The price per million input and output tokens (the chunks of text a model builds its answers from) stays at 2 and 10 dollars (Anthropic).

On Terminal-Bench 4.0, which measures command-line work, it scores 70.6% according to the company, compared with 10.3% for Sonnet 5.
It is the first Sonnet to have the same cyber safeguards as the more expensive Opus models. It also has protection against distillation, that is, against competitors training their own models on its answers.
The model is available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. The older Sonnet 4.5 will be retired on November 30, 2026 (Claude Platform).

Google Unveils Gemini 4 Argon (September 30)

Google released Gemini 4 Argon and called it its most capable model to date. It focuses on programming, research and cybersecurity. According to the company, it can independently find, verify and fix critical vulnerabilities in software (TechCrunch).

For now, only a select group of security partners in the Fairwind program has access. A broad release will come only after safety testing that Google is conducting with the US government. Google says Argon outperforms OpenAI’s GPT-6 Astra and Anthropic’s models on a range of benchmarks; these are, however, the company’s own claims (CNBC).

Products and Agents

OpenAI DevDay: Dots Agents and New Plans (September 29)

At its DevDay developer conference in San Francisco, OpenAI showed more than 20 announcements to roughly 2,500 attendees. The headliner was Dots, “always-on” AI agents, meaning programs that carry out tasks on their own without constant prompting. Each agent has its own cloud computer, connects to more than 4,000 apps and can be messaged in ChatGPT, Slack or Teams (CNBC).

ChatGPT Space is a shared space where team members and Dots agents work together. Pages is a new document type that both people and agents write in.
Ultrafast is a paid speed tier. According to CNBC, it speeds up text generation up to eight times in Codex and up to six times in the API.
The new Pro 500 plan offers the highest usage limits and access to Ultrafast.
OpenAI Private Intelligence is still in preview and promises businesses better control over their data. Cisco, Databricks and Snowflake were named as partners.
The company also announced the GPT-6.1 Sol model, which arrived a week after GPT-6 Sol.

Dots agents are for now available only to ChatGPT Pro and Business Premium subscribers in selected countries. CFO Sarah Friar said the enterprise business has doubled since July.

Business and Investment

Barclays Rolls Out Claude Across the Bank (October 1)

British bank Barclays is expanding its partnership with Anthropic. By the end of 2026 it wants half of its developers using Claude Code, and most of them in 2027. More than 16,000 employees already use an internal assistant built on Claude (Anthropic).

In the Global Markets division, Claude models sort and route about 120,000 client emails every day.

Anthropic Commits 100 Million Dollars to Training Engineers (October 2)

The Claude Frontier Academy program aims to train 10,000 so-called Frontier Deployed Engineers by the end of 2027. These are specialists who deploy AI directly inside companies. The first cohorts include people from Accenture, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk (Anthropic).

More Business News

September 29 According to CNBC, OpenAI is in early talks about further funding of roughly 30 billion dollars. Altman said he has no fixed date for going public on the stock market (CNBC).
September 30 Voice company ElevenLabs completed a 300 million dollar employee share tender. This doubled its valuation since February to 22 billion dollars. It also released the text-to-speech models Eleven v4 and v4 Turbo (Tech.eu).
September 30 China’s DeepSeek open-sourced tools for Huawei Ascend chips, including the TileLang programming language. The goal is to depend less on Nvidia chips (The Standard / Reuters).
September 30 Nvidia announced that cloud company CoreWeave is deploying Vera Rubin NVL72 systems for AI agents and named Cognition as a customer (Nvidia).

Regulation and Law

A Voluntary Safety Commitment at the White House (September 29)

President Donald Trump invited tech company leaders to lunch at the White House. There they signed a one-page document, the Joint Commitment on Frontier Responsibilities. It was signed by Dario Amodei (Anthropic), Greg Brockman (OpenAI), Sundar Pichai (Google), Mark Zuckerberg (Meta), Elon Musk (xAI) and Jensen Huang (Nvidia) (Fortune).

The companies promise four layers of oversight: internal testing, audits by an outside firm, board-level review and regular meetings on shared standards. They mainly want to monitor risks in cybersecurity and in biological and chemical threats. Trump called the commitment “morally binding”, but it carries no penalties (CBS News).

“Artificial Intelligence” Becomes “Super Intelligence” (September 29)

On the same day, Trump signed an executive order titled “Inaugurating the Era of Super Intelligence”. Federal agencies are to use “Super Intelligence” (SI) instead of “Artificial Intelligence” (AI) in their communications and documents. The legal meaning of the term does not change, and the president’s science adviser has 60 days to propose a statutory definition (The White House).

FTC Investigates OpenAI and Anthropic (September 30)

The day after the commitment was signed, it emerged that the US Federal Trade Commission (FTC) is investigating OpenAI, Anthropic and other AI labs. It is examining whether they violated the FTC Act, which protects consumers. The focus is mainly on the risks of AI agents following incidents such as this year’s breach of the Hugging Face platform’s systems (The Next Web).

The agency is preparing so-called civil investigative demands, official requests similar to subpoenas that can even compel testimony from executives.
According to an FTC spokesperson, this is only an investigation for now, not a lawsuit. No company has been accused of breaking the law.

New Government “Super Intelligence Force” (October 4)

Trump announced the creation of the Super Intelligence Force, a task force meant to coordinate federal AI policy. It is led by Director of National Intelligence Jay Clayton. Its members also include FTC Chairman Andrew Ferguson and Pentagon chief technology officer Emil Michael (Spectrum News / AP).

Courts: AI Training and Google Search

September 29 The US Court of Appeals for the Third Circuit upheld Thomson Reuters’ win over startup Ross Intelligence. Ross trained its legal search engine on headnotes from the Westlaw database, and according to the court this is not “fair use” (the legal exception for permitted use of a work). It is the first such ruling by a US appeals court in disputes over AI training. However, Ross was not building generative AI, so the impact on large language models will be limited (Nieman Lab).
October 1 Judge Amit Mehta dismissed antitrust suits brought by Chegg and Penske Media against Google. The companies argued that AI Overviews summaries in search were taking away their traffic. The court, however, found no violation of the Sherman Act (Northeast Times / AP).
September 29 The European Commission opened a public consultation on copyright and AI. It covers the use of protected content in model training, online piracy and research, and runs until November 3, 2026 (European Commission).

Safety and Incidents

Safety ran through the entire week. The withdrawal of GPT-6.1 Astra showed that even a major lab will shelve a finished model when it behaves unreliably in testing. The main issue was that the model inaccurately described to users what it had done (CNN / KRDO).

Google also opted for a gradual rollout of Gemini 4 Argon. According to CNBC, it wants to strengthen protections in four areas before the public launch, for example against misuse and against prompt injection. That is an attack in which planted text manipulates a model into doing something other than what the user wants (CNBC).

Anthropic’s Sonnet 5.5 is the first in its class to have cyber safeguards that previously only the more expensive Opus models had (Anthropic).
According to TNW, the FTC investigation focuses mainly on the risks of AI agents following recent incidents, including the breach of Hugging Face (The Next Web).