Google holds back Gemini 4 Argon; AI chiefs sign Trump's accord

Google DeepMind announced Gemini 4 Argon, its first new frontier flagship since Gemini 3.1 Pro in February, after a promised Gemini 3.5 Pro never shipped. For now it is going only to vetted security teams in the Fairwind Program, the gate Google set up in early September for its Gemini 3.8 Flash Cyber model, while safeguards are hardened. Google says Argon leads the DeepSWE software engineering benchmark at 77.9 percent and tops the Vals Index of finance, legal and tax work, and it raises the output limit to 1 million tokens, up from 64,000. A monitor watches the model's chain of thought and stops runs that go out of bounds. Paid API customers and Google AI Ultra subscribers come next, with no date given, at an introductory $2 per million input tokens and $10 per million output, rising to $4 and $20 afterward.

The full text of the Joint Commitment on Frontier Responsibilities, shared by White House adviser David Sacks, shows what President Trump and the leaders of Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed on Tuesday, alongside roughly two dozen firms in all. Each company agrees to monitor its models for cyber, biological and chemical risks and for reaching systems they were never meant to touch, to empower an internal team to fix problems, to hire an independent auditor to test those controls, and to have a board committee oversee the results. Nothing is legally required and no penalty is attached, though Trump called the deal "morally binding." The text says it may later make sense to write the steps into law. Google, Anthropic and OpenAI had already committed to outside audits, The Information reported.

OpenAI says it disrupted a campaign that began on July 1 to pull out its models' protected reasoning, the hidden working behind a final answer, and attributes a core cluster of the activity to individuals associated with Moonshot AI, the maker of Kimi. One technique copied encrypted reasoning from one conversation and asked a model in another to decrypt and transcribe it. Attempts spiked on July 24 and 25, with 16,000 requests from more than 4,000 accounts, and OpenAI says it had shut down a wider cluster of over 15,000 users by July 28. It closed the replay path and shared its findings through the Frontier Model Forum. The disclosure follows Anthropic's distillation report on September 10, which also named Moonshot.

On his last day to act on bills, California Governor Gavin Newsom signed laws that bar employers from relying on AI to decide to fire someone, ban using biometric data to infer a worker's emotional state, and require written notice when AI is responsible for mass layoffs. He vetoed three other AI bills, covering health care services, notice to public employees when AI performs their work, and AI used by mental health professionals. Unions had pressed for a slate of eight bills, and the California Labor Federation said work remains on health care. Newsom also ordered state agencies to keep saying "artificial intelligence" after President Trump's rename, and left open a special session on AI.

Legal Advocates for Safe Science and Technology, a nonprofit, sued OpenAI in San Francisco County Superior Court, arguing that its agents' July break-in at Hugging Face violated California's computer-fraud law, and that state law says it is no defense that an AI caused the harm on its own. The group wants an order barring OpenAI's agents from entering other systems without permission and halting what it calls unsafe development practices, and it seeks no damages beyond attorneys' fees. OpenAI called the suit completely without merit, pointing to its technical report and the model it held back. The filing lands days after Florida asked a state court to pause OpenAI's frontier development.

Google DeepMind introduced SynthID Bio, which hides a detectable signature in the amino-acid choices of an AI-designed protein. In wet-lab tests on binders for three targets, among them the SARS-CoV-2 spike protein and the cancer target PD-L1, watermarked designs matched the hit rate and binding strength of unwatermarked ones. A version built into AlphaFold 3 marks predicted structures too. The aim is DNA synthesis screening: AI can produce sequences unlike any known threat, so a watermark lets screeners recognize designs from trusted models and focus scrutiny on the rest. DeepMind calls it a proof of concept.

Google has admitted about 100 publishers to a pilot that pays them when their content materially contributes to AI Overviews and other Gemini-powered search answers. Several small and mid-sized sites told The Information the payments come to roughly one-tenth of one percent of their advertising revenue. One early participant is on track for more than $1 million a year, while several smaller sites earned less than $1,000 over several months. Some larger publishers have declined to join in hopes of a better deal, and publishers say it is unclear what counts as a meaningful contribution.

ElevenLabs let employees sell $300 million of vested shares in a tender offer valuing the voice AI company at $22 billion, double the $11 billion it reached when it raised $500 million in February. Wellington and T. Rowe Price co-led the purchase, institutional investors that typically keep their stock after a company goes public. It is the company's second employee sale, after one at $6.6 billion in September 2025.