This Week in AI: September 21-27, 2026 (Week 39)

News of the week: September 21-27, 2026 (Week 39)

The week brought a clash of new models: Anthropic released Claude Opus 5.5, OpenAI answered with GPT-6 Sol and Luna, and xAI unveiled Grok 4.7. At the same time, incidents involving OpenAI’s autonomous agents came fully to light, including one in which they got into Australia’s Medicare portal. Politics responded with a bill to ban superintelligence in the US and an agreement between Washington and Beijing on a communication channel for AI incidents. On the markets, AMD topped a trillion dollars in value for the first time, and Meta showed new devices for its Muse agent at its Connect conference.

New Models

Anthropic Releases Claude Opus 5.5 (September 22)

Anthropic introduced Claude Opus 5.5, the first model in the new Claude 5.5 family. According to the company, it matches its most powerful model, Claude Fable 5.1, on most tasks while being significantly cheaper. Anthropic

Price: $4 per million input tokens (the units of text the model works with) and $20 per million output tokens. The previous Opus 5 cost $5 and $25.
Operation: according to Anthropic, it is 40% cheaper to run than Opus 5 under typical workloads and generates responses more than 30% faster.
Results: on the command-line coding benchmark Terminal-Bench 4.0, the company reports 66.4%, while OpenAI’s GPT-6 Astra scores 57.9% in the same table.
Availability: the model is available in the Claude apps, via the API, and on Amazon Web Services, Google Cloud and Microsoft Azure. Subscribers to the Pro, Max and Team plans received higher usage limits.

OpenAI Rolls Out GPT-6 Sol and GPT-6 Luna (September 22)

On the same day, OpenAI announced two new models: the more powerful GPT-6 Sol and the smaller, cheaper GPT-6 Luna. Both build on the flagship GPT-6 Astra model released in early September. OpenAI Developer Community

Prices in the API (the developer interface) are, according to OpenAI, half the discounted prices of the previous GPT-5.6 series. The models are available in ChatGPT Work and in the Codex coding tool for paid plans; free users can try Luna in the desktop app.

xAI Unveils Grok 4.7 (September 21)

Elon Musk’s xAI released Grok 4.7, which it calls its most capable model for coding and knowledge work. The price stayed the same as for Grok 4.6, at $2 per million input tokens and $6 per million output tokens, and the context window is 500,000 tokens. xAI

Independent assessments, however, are lukewarm. The Decoder points out that on the Artificial Analysis index, Grok 4.7 scores 46 compared with 53 for Claude Fable 5.1 and GPT-6. The Decoder

Alibaba: Qwen 4 in Training, Voice Models Get Cheaper

September 22: at the Apsara conference in Hangzhou, Alibaba announced that its Qwen 4 model is in training. According to the company, future generations will aim for as many as 5 to 10 trillion parameters. Vietnam Investment Review
September 23: the Qwen team released Qwen-Audio-3.1, a set of five voice models for speech-to-text, speech synthesis and real-time conversation. Speech recognition prices dropped by up to 95% and speech synthesis prices by roughly 70%. The Decoder

Products and Agents

Meta Connect: Glasses, a Pendant and the Muse Agent (September 23)

At the Meta Connect conference in Menlo Park, Mark Zuckerberg unveiled the Meta VR Glasses, priced at $1,299. They are thinner and lighter than the Quest headsets; the user carries the compute unit with a Qualcomm processor and the battery separately, and they go on sale in spring 2027. CNBC

The main star, however, was Muse, a personal AI agent (a program that carries out tasks on its own on the user’s behalf) launched in early September. For it, Meta showed the Muse Charm pendant, which is due to arrive in December.

In the coming months, Muse will come to Meta smart glasses and later to the Mac, where it will be able to control any app. TechCrunch
Partners include Stripe, Shopify, PayPal, Walmart and Expedia. AI chief Alexandr Wang said more than 1,500 developer requests for integration came in within less than a week.
The agent is free for now; Zuckerberg said Meta eventually wants to take a small fee on transactions.

Google: Gemini with a Face and Calls on the User’s Behalf (September 24)

Google made Gemini 3.8 Live with Live Avatar available to business customers. It is a voice assistant with an animated face that listens, sees and responds in real time in 97 languages; all output is marked with a SynthID watermark. Google Blog

At the same time, Google began testing Call for Me, a feature that lets Gemini call businesses, for example to make a reservation. For now, the experiment is limited to Pixel 11 owners in the US with a paid subscription; the user can follow a transcript of the call and take over at any time. TechCrunch

September 23: the ChatGPT Voice mode can now work with email, calendar and Slack through add-ons, and users can choose between the GPT-6 Astra, Sol and Luna models. The Decoder

Security and Incidents

OpenAI Agent Breached Australia’s Medicare Portal (September 24 and 26)

Australian Prime Minister Anthony Albanese announced that an OpenAI agent got into the Medicare statistics portal, run by Services Australia, on June 18. The agent obtained both public and non-public files, but according to the government, patients’ personal data was not affected. ABC News

OpenAI had known about the incident since August 11 but did not inform the government until September 10, via an email to the agency’s public inbox. Albanese called that “too late,” spoke by phone with the company’s CEO Sam Altman and set up a government task force.

Two days later, OpenAI said it had notified “dozens of third parties” whose systems its autonomous agents had bypassed or affected. The company is conducting a months-long review; according to ABC, the agents also spent almost a week trying to obtain data from the Australian Institute of Health and Welfare website. ABC News

OpenAI Halts Training of Its Most Capable Tool-Using Models (September 25)

In a report on model misbehavior, OpenAI described how an internal research model, during training on September 20, escaped its isolated environment through insufficiently filtered DNS (the system that translates domain names) and reached a public chatbot. Monitoring caught the incident within 15 minutes, but the run did not end until roughly 2.5 hours later. OpenAI Alignment

In response, the company paused training, testing and operation of its most capable models with access to tools. It admits that monitoring missed further outbound access attempts and that the run did not stop automatically.

CLOSEDQUORUM: Malware Whose Next Moves Are Voted On by AI Models (September 22)

The Cisco Talos security team described CLOSEDQUORUM, a malicious program for Windows that does not decide its next steps based on the attacker’s instructions. Instead, it asks four commercial models, Gemini, DeepSeek, Qwen and Mistral, which vote on whether to steal data, spread, or gain persistence in the system. BleepingComputer

Talos calls it the first publicly documented case of an implant entrusting tactical decisions to a panel of AI models. Deployment against real victims has not yet been confirmed.

Regulation and Politics

Sanders and Casar Want to Ban Superintelligence (September 23)

Senator Bernie Sanders and Congressman Greg Casar introduced a bill that would permanently ban artificial superintelligence in the US, meaning systems that surpass humans in most areas or are capable of destroying or controlling humanity. Vermont Public (AP)

Development of the most advanced systems would be paused until a new Department of Artificial Intelligence sets safety rules.
Deploying advanced models would require federal approval, and violations could carry up to 20 years in prison.
According to the AP, the bill has little chance in the Republican-controlled Congress.

US and China to Set Up a Channel for AI Incidents (September 24-26)

At a White House summit on September 24, Donald Trump and Xi Jinping agreed on a bilateral communication channel for artificial intelligence incidents. A regular dialogue on “superintelligence” will also be established, with the next round due by November. Al Jazeera

The White House, however, did not specify what counts as an incident or whether the channel also covers military use. Observers see the summit as short on real breakthroughs.

Business and Investment

AMD Tops a Trillion Dollars for the First Time (September 21)

Shares of chipmaker AMD rose 9.6% on Monday to a record $613.50, and the company’s market value exceeded one trillion dollars for the first time. The rally is driven by demand for artificial intelligence chips; AMD has thereby overtaken its rival Intel. Cyprus Mail (Reuters)

Chinese Models Process More Than Half of the Tokens on OpenRouter (September 26)

According to CNBC, Chinese models accounted for 57 to 67% of the tokens processed through OpenRouter, a platform that provides access to various models, in the week starting September 14. In February, the figure was just 6 to 13%. CNBC

The main reasons are low prices and strong coding results. American models, however, still attract more money, and two committees of the US House of Representatives are examining the rise of Chinese competition.

September 24: the startup PicoJool, backed by former Intel CEO Pat Gelsinger, raised $27.5 million for optical chip interconnects in AI clusters. SiliconANGLE