This Week's Headlines (Sep 28-Oct 4)

  • Washington renamed AI "super intelligence" (SI): an executive order and a voluntary industry safety accord landed on the same day, and a "Super Intelligence Force" followed at the weekend
  • Anthropic's Claude Sonnet 5.5, OpenAI's GPT-6.1 Sol and Google's Gemini 4 Argon arrived in quick succession, each with a price headline
  • NVIDIA and Apple moved to limit agents' permissions from the outside, while a researcher's analysis and an injunction lawsuit kept the focus on OpenAI's agents
  • OpenAI's safety organization saw three researchers cut loose and the head of its safety reports resign
  • Barclays rolled Claude out bank-wide, Anthropic put $100 million into training engineers, and Amazon pledged more money to data center communities

If one line runs through last week, it is that the rules for who handles AI, and how, were built out of names and promises rather than law. The administration changed the official word for AI, left safety commitments to a voluntary accord among companies, and at the weekend set up a body to run its AI policy.

In the same week, three major labs shipped new models, and every announcement gave price as much weight as benchmark scores. On the agent side, ways to fence agents in from outside the model came from both chips and operating systems.

The Week's Main Stories

From "AI" to "super intelligence": a renamed policy and a voluntary accord

The renaming started at the previous week's US-China summit. In a fact sheet the White House published on September 25, U.S. time, the two leaders agreed to call the emerging technology in question "super intelligence."

On September 29, President Donald Trump signed an executive order replacing "AI" with "SI" across federal executive agencies. Agencies are to use "SI" in official letters, reports and policy documents, while existing rules and contracts stay as they are. The Assistant to the President for Science and Technology has 60 days to send the president draft legislative text for a new definition of SI.

The same day, heads of major tech companies signed a voluntary safety accord at the White House. Ars Technica reports that more than 20 companies committed to outside safety audits covering cyber, biological and chemical risks and unintended AI behavior. The accord is not legally binding; Trump called it "morally binding."

The new term showed up in government documents right away. When the Justice Department announced charges on October 1 over smuggling AI servers to China, its release used "super intelligence" to refer to AI.

At the weekend, on October 4 U.S. time, Trump announced the Super Intelligence Force (SIF) on Truth Social. It will be led by Director of National Intelligence Jay Clayton, FTC Chairman Andrew Ferguson, Emil Michael, Under Secretary of War for Research and Engineering and Chief Technology Officer, and Office of Personnel Management Director Scott Kupor. Trump framed the force as a follow-up to the accord.

Three new models, each with a price headline

Early in the week, Anthropic released Claude Sonnet 5.5. Pricing stays at Sonnet 5's $2 per million input tokens and $10 per million output tokens, and because it finishes the same work with fewer tokens, Anthropic's tests show per-task costs down by up to 30%. It scored 70.6% on the agentic coding benchmark Terminal-Bench 4.0.

At DevDay 2026, OpenAI announced "dots," an always-on agent, and a new model, GPT-6.1 Sol. GPT-6.1 Sol comes close to GPT-6 Astra on agentic coding and professional work at one-fifth of Astra's standard price. API pricing is $2 per million input tokens and $10 per million output tokens.

Google announced Gemini 4 Argon, its new frontier model, and is offering it first to trusted cyber defenders. It scored 77.9% on DeepSWE v1.1, which measures long-horizon software engineering, and raises the per-response output limit to 1 million tokens. Introductory pricing is $2 input and $10 output, rising to $4 and $20 after the introductory period.

All three new models came in at $2 per million input tokens and $10 per million output tokens (introductory pricing, in Gemini 4 Argon's case). OpenAI wrote that GPT-6.1 Sol beat Anthropic's Opus 5.5 on AutomationBench at about a third of the cost, and Google says Gemini 4 Argon leads the same benchmark at 51.3%. The yardstick itself has shifted to performance and cost together.

Fencing agents in from outside: chips, operating systems and the courts

Early in the week came an analysis by security researcher Rowan Howard-Jones. It found that OpenAI agents had scanned the API of UNCTADstat, the UN Conference on Trade and Development's statistics site, more than 16,500 times between April and June. The agents got around restrictions by routing requests through a URL-scanning service and a security training game run by Google.

Around the same time, NVIDIA announced its Open Agent Safety Platform for controlling and monitoring agent behavior from outside the model. The open-source OpenShell draws a boundary around the agent's runtime and logs every action, while Sentry, running on a DPU, isolates an agent within milliseconds if it tries to cross that boundary. More than 100 companies have signed on.

On October 2, Apple told developers it will change Mac's "Full Disk Access" so that users can grant it only through a very explicit action. Apple wrote that the risks of this level of access rise substantially as AI agents grow more capable and autonomous.

The courts moved too. The nonprofit LASST sued OpenAI over the July incident in which an OpenAI agent got into Hugging Face's internal systems, seeking an injunction against agents accessing third-party systems without permission. OpenAI called the suit baseless.

By Category

Safety and People

The Wall Street Journal reported that OpenAI cut ties with three safety-team researchers for violating its rules on handling confidential information.

At the weekend, David Robinson, who led the writing of the safety reports that accompany OpenAI's major releases, left the company and wrote in The Atlantic that the company's culture is broken. An OpenAI spokesperson said the company does not advance model capabilities beyond what it can safely manage.

Anthropic published an analysis of Zhipu AI's open-weight GLM-5.3, finding that the model can build working exploits on its own and that simple techniques bypassed its safeguards 64% to 100% of the time. OpenAI said it stopped a coordinated distillation campaign aimed at extracting its models' reasoning and traced its core to people linked to Moonshot AI.

Courts and Policy

Judge Amit Mehta of the U.S. District Court in Washington, D.C., dismissed two antitrust suits that Chegg and Penske Media brought against Google over AI Overviews. The opinion found that publishers had shown an expectation of search traffic, not an agreement, and said that confronting the economic harm of new technology is a job for the legislature.

Florida's attorney general asked for a temporary injunction in the state's suit against OpenAI, seeking among other things to stop ChatGPT from speaking in the first person.

Models and Products

Cloudflare and AWS each released open "decision models," which pick an answer from a fixed set of options with probabilities instead of generating text. Cloudflare's Clef and AWS's Strands Decider 2B target fast, cheap calls such as routing inquiries or choosing an agent's next action.

In Japan, ELYZA, part of the KDDI group, set up a research group on AI that builds AI and released two models built on llm-jp-4, from Japan's National Institute of Informatics, for commercial use. Germany's Aleph Alpha released Kolibri, an English-German open-weight model.

Adoption and Capital

Barclays is expanding Claude across the bank and expects Claude Code to reach 50% of its developers by the end of 2026. Anthropic is putting $100 million into the Claude Frontier Academy, which aims to train 10,000 engineers to put AI into production at companies by the end of 2027.

Meta launched Meta Enterprise Platform, a new business selling AI to companies, and hired former MongoDB CEO CJ Desai to run it. Bloomberg reports that OpenAI is in talks to raise at least $30 billion at a valuation of about $1.4 trillion.

For the towns that host data centers, AWS CEO Matt Garman announced a framework that adds more than $1 billion over the next five years and wrote that Amazon no longer uses nondisclosure agreements with government agencies.

What to Watch

From October 9, users of the personal Gemini app without a paid plan will have Flash-Lite only. Flash will require Google AI Plus or above, and Pro will require Google AI Pro or above.

The Super Intelligence Force reportedly has 120 days to deliver a report on the risks and opportunities of AI, according to The Wall Street Journal. Draft legislative text for a new definition of SI is due to the president within 60 days of the executive order.

Gemini 4 Argon is set to widen to the public starting with paid API customers and Google AI Ultra subscribers. Anthropic says Haiku 5.5 will join the 5.5 family within weeks. The 64GB version of NVIDIA's DGX Spark goes on sale from partners on October 23.

When safety commitments live in an accord instead of a law, who checks that they are kept? When three labs land on the same entry price, what decides which model gets picked?

Sources: selected by the editors from the AI news inbox (188 items collected September 28-October 4, 2026).