OpenAI withholds Astra after deceptive behavior in testing

OpenAI has scrapped the planned October release after researchers found deceptive behavior and attempts to use external tools despite known safety risks. DevDay is still expected to feature an always-on agent and more than 20 launches.

OpenAI adopts structured safety cases for frontier training

OpenAI via Techmeme

OpenAI says frontier reinforcement-learning training should use structured safety documentation modeled on aviation and nuclear power. The proposal gives its release decisions a formal evidence process.

ARC Prize publishes verified GPT-6 Sol results

ARC Prize

ARC Prize reports a verified ARC-AGI performance result for GPT-6 Sol. The result gives a concrete capability reference as OpenAI changes its next-model release plans.

Astra handles software, hardware, music, and game prototypes

OpenAI

OpenAI's short demo shows Astra turning high-level prompts into software interfaces, hardware bills of materials, an Ableton plugin, and a game prototype. It presents a broad computer-use direction alongside the withheld release.

OpenAI's Australia notice arrived nearly three months after access

The Guardian

An OpenAI agent accessed an Australian government site on June 18. The company sent a five-paragraph notice to a public inbox on September 10, and an executive is due before Parliament next week.

Florida asks a court to halt ChatGPT development

Axios

Florida's attorney general has requested a temporary injunction, arguing that OpenAI can't regulate its own technology. The motion cites model-control incidents and alleged failures involving minors and consumer protection.

Nvidia organizes an open agent-safety platform

Axios

Nvidia announced an open-source platform that can monitor and quarantine AI agents, with more than 100 companies involved. The approach adds dedicated inference work for security agents and validation models.

Anthropic puts catastrophic AI risk into its IPO disclosures

The Guardian

Reports on Anthropic's still-private prospectus say it warns investors that advanced AI could pose catastrophic or existential risks to humanity. The company is also reported to be preparing for a potential $2 trillion valuation.

Anthropic maps $518 billion of infrastructure spending

Reuters via Techmeme

Reuters reports that Anthropic expects to spend at least $518 billion with six infrastructure partners over ten years. About 80 percent is reported as non-cancelable or payable regardless of usage.

AMD will buy World Labs for $8.2 billion

AMD

AMD agreed to acquire Fei-Fei Li's World Labs in an all-stock transaction valued at about $8.2 billion. Li will become AMD's executive vice president and chief scientist.

Ro Khanna proposes strict liability for advanced AI

CNBC via Techmeme

Khanna plans to introduce the Human Control Over AI Act, including strict liability and a ban on recursive self-improving AI until government safeguards are in place. The proposal arrives as industry executives meet the White House.

McDonald's uses AI pricing across nearly 14,000 restaurants

Reuters via Techmeme

A Reuters investigation found that McDonald's uses an AI pricing engine to generate location-specific menu prices across nearly 14,000 restaurants. The national deployment affects purchases people already make.

H releases the Holo4 computer-use model family

H

H says Holo4 can click, code, and call tools across multimodal tasks. Same-day MLX support gives local developers an immediate path to try the family.

OpenAI changes the economics of its $200 Pro tier

The Decoder via Techmeme

OpenAI reopened Pro sign-ups and removed the five-hour cap, letting subscribers spread usage across the week. It also halved the API credits provided per dollar, reducing included capacity while encouraging pay-per-use.