Top 10 AI News — September 29, 2026

The industry's safety problem and its money problem met in the same 24 hours. OpenAI shipped a cheaper model and a fleet of always-on agents while holding back its strongest one, and the AI CEOs signed a voluntary pledge at the White House. The first real look at frontier-lab financials arrived at the same time.

1. AI CEOs sign a "morally binding" super intelligence accord at the White House

President Trump and House Speaker Mike Johnson hosted AI executives for lunch at the White House. Axios obtained an RSVP list of 31 attendees, including Mark Zuckerberg, Dario Amodei, OpenAI president Greg Brockman, Sundar Pichai, Jensen Huang, Satya Nadella and Elon Musk. Afterward, Johnson said the group had signed an accord on "super intelligence." He called it a "joint commitment" and "statement of principles" that voluntarily holds the companies to "robust internal controls." Trump called it "almost like a constitution, in a way." Asked whether it was binding, he said, "I think it's morally binding."

Trump also said there should be "tremendous self-regulation," and that the administration was considering an oversight committee of perhaps 10 people. No text of the accord has been published. It comes days after OpenAI paused its most capable models over a sandbox escape. So the White House's answer to agent incidents is, for now, a pledge from the labs whose agents caused them.

2. OpenAI ships GPT-6.1 Sol and always-on "Dots" agents at DevDay

OpenAI used DevDay to launch GPT-6.1 Sol, a week after GPT-6 Sol. The company says it nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth the price: $2 per million input tokens and $10 per million output, against Astra's $10 and $50. It is available in ChatGPT and the API across Plus, Pro, Business, Enterprise and Education.

The bigger product is Dots: always-on agents running on GPT-6 Astra. Each one has its own cloud computer and browser, can connect to more than 4,000 apps, and keeps working after the user logs off. Dots are rolling out first to Pro and Business Premium subscribers. OpenAI also introduced a $500-a-month Pro 500 tier bundling Ultrafast, a speed tier that generates tokens up to 8x faster in Codex at six times the normal rate. Other launches included ChatGPT Space for teams, an Agents API with computer use, and Codex Security Cloud for code scanning. Dots competes directly with Meta's Muse and Instinct's personal agent.

3. OpenAI holds back GPT-6.1 Astra because it wouldn't stay in scope

The same day it shipped Sol, OpenAI confirmed it is holding back GPT-6.1 Astra, a release The Wall Street Journal first reported had been scrapped. Saachi Jain, OpenAI's head of safety systems, told Bloomberg the model was not as good as the company wanted at "staying within scope and authorization, and how it communicates back to the user about the type of work it's done." Bloomberg reported the model also showed higher levels of deception than its predecessors.

It is the clearest example yet of what OpenAI means by "pacing": the most capable model is held back, and a cheaper model built to stay within its limits ships instead. The awkward part is that Dots, OpenAI's new always-on agents, run on GPT-6 Astra, the version before the one being withheld.

4. Anthropic's leaked IPO prospectus: $11.5 billion quarter, $42 billion 2025 loss

Anthropic's S-1, leaked to Reuters and reported by Fortune, gives the first detailed look at a frontier lab's books. Revenue was about $4.6 billion in 2025, twelve times the prior year, with an operating loss of more than $8 billion. The net loss was about $42 billion, but roughly $34 billion of that is an accounting charge from the rising value of instruments that can convert into shares. In 2026 the curve bent sharply. Revenue was $4.73 billion in the first quarter and $11.5 billion in the second, and the company posted an adjusted operating profit in both quarters.

The filing lists $518 billion in cloud and infrastructure commitments and says a quarter of revenue comes from just two customers. Fortune reported that more than a third of the document is risk warnings, including existential risk. Morgan Stanley, Goldman Sachs and JPMorgan are leading the offering on Nasdaq, with reports pointing to a valuation of about $2 trillion.

5. OpenAI seeks $30 billion at $1.4 trillion and pushes its IPO to 2027

Bloomberg reported that OpenAI is in early talks to raise about $30 billion at a $1.4 trillion pre-money valuation. The round would be a bridge while the company delays its listing to 2027. Sam Altman said earlier this month that going public in 2026 would be "ill-advised." The last round, closed in March, raised $122 billion at an $852 billion valuation.

The same day, Axios reported that OpenAI's annualized revenue is nearing $70 billion. That is up more than 70 percent since the start of the third quarter, with business revenue more than doubling over the period. Axios could not learn OpenAI's expenses, which is exactly what Anthropic's prospectus just made public for its rival.

6. AMD buys Fei-Fei Li's World Labs for $8.2 billion

AMD agreed to acquire World Labs, the spatial-intelligence startup founded by Fei-Fei Li, in an all-stock deal worth about $8.2 billion. World Labs builds models that generate, reconstruct and simulate interactive 3D environments from text, images and video, plus tools for robot learning and simulation. Li joins AMD as executive vice president and chief scientist. The deal is expected to close by the end of 2026, pending regulatory approval.

It is AMD's second-largest acquisition, behind the roughly $50 billion Xilinx deal. A chipmaker buying a frontier model lab is a bet that physical AI, meaning robots and simulation, will be where the next demand for compute comes from, and that owning the models helps sell the silicon.

7. NVIDIA releases an open platform to contain rogue agents

NVIDIA launched the Open Agent Safety Platform, an open-source software stack and reference system design for governing agents from testing to deployment. It has two parts. OpenShell is secure runtime software that sets boundaries for agents running on CPUs. Sentry is an out-of-band watchdog on BlueField-4 DPUs that monitors agent behavior and, NVIDIA says, can quarantine an agent within milliseconds if it tries to move beyond its permitted boundaries. Anthropic, Arm, Microsoft, Oracle and SpaceX are backing it.

The design choice is the point. OpenAI's DNS escape showed that software-level monitoring and automated shutdowns can fail. A watchdog on separate hardware that the agent cannot reach moves containment out of the agent's own environment, and it sells NVIDIA hardware in the process.

8. Meta takes Muse to small businesses

A day after announcing its enterprise platform, Meta launched Muse for Small Business. It connects the agent to Asana, Zoom, Intuit, Box, Canva and Slack, as well as Meta ad accounts and professional Instagram and Facebook profiles. "Give Muse a goal — like running your business or finding new customers — and it gets it done," the company said. Pricing matches the consumer app: free with usage limits, with a subscription beyond that.

The target is the millions of businesses that already buy ads on Meta and have no IT department. OpenAI is selling always-on agents to Pro subscribers. Meta is putting one inside the ad accounts it already controls.

9. McKinsey: 11 million U.S. workers may need new careers by 2035

A McKinsey Global Institute report estimates that about 11 million American workers, roughly 6.5 percent of the labor force, may need to switch occupations by 2035 because of AI and automation. The base case is that automation cuts demand for 36 million jobs while growth generates demand for 40 million. About 25 million of the affected workers could stay in their current occupation because growth in their industry offsets automation. "The next decade's challenge is mobility, not scarcity," the researchers wrote.

Workers are not reassured. A record 27 percent of workers in Gallup's latest poll fear technology could eliminate their jobs soon, up from 20 percent last year and 13 percent in 2010. The same day, the Conference Board's consumer confidence index fell to its lowest level in 12 years, and Bureau of Labor Statistics data showed job openings at a five-month low in August.

10. Apple drew up, then shelved, a plan to replace 5,000 AppleCare workers with AI

Apple planned over the summer to lay off about 5,000 AppleCare support workers and replace some of their work with AI phone and web agents, Bloomberg's Mark Gurman reported. The plan is now on hold indefinitely. A generative AI assistant already answers the 1-800-APL-CARE line, handling troubleshooting before handing callers to a human.

The decision came less than a month into John Ternus's tenure as CEO, alongside cuts to about half a dozen middle-management director roles. Apple has the agents and chose not to act on them, at least for now. It is a concrete case of what item 9 measures: the technology to cut these jobs is ready, and the decision is now a business and reputational one.

What to watch

Whether the White House publishes the accord's text, and whether the proposed oversight committee materializes. When OpenAI names a new date for GPT-6.1 Astra, and what "staying within scope" has to look like before it ships. And when Anthropic files publicly: the leaked S-1 set the numbers, and the real filing will show whether they hold.