Google holds back Gemini 4 Argon as Washington sticks to voluntary AI rules

The last few days of public AI news turned on who may use the most capable systems, who pays for the computers behind them, and how little force new safety promises carry.

Google limits Gemini 4 Argon

On 30 September, Google DeepMind announced Gemini 4 Argon for long, multi-step work in software engineering, legal and finance tasks, and cybersecurity defense. Koray Kavukcuoglu, senior vice president of Google DeepMind, wrote that it is rolling out first to trusted cyber defenders through the Fairwind Program. Google said it is in the U.S. government’s voluntary process for pre-release access and will tighten guardrails before a wider release, starting with paid API customers and Google AI Ultra subscribers. It gave no date.

The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input discounted 95 percent. Later rates would be $4 and $20. The output limit rises from 64,000 tokens to 1 million.

Google’s own scores, not independently confirmed here, include 77.9 percent on DeepSWE v1.1, a tie for first on CWE-bench v1 at 68 percent, 91.7 percent on LVBench, and 51.3 percent on Zapier’s AutomationBench. Trusted defenders and internal teams will get a version without cyber guardrails. Reuters, citing a spokesperson, reported that Argon is larger than Google’s previous Pro models and that Google sees it as comparable to OpenAI’s Astra and Anthropic’s Opus on key coding and cyber benchmarks. On Google’s charts, Reuters noted, Argon still trailed on some tests, including two of four coding benchmarks in the release. Google also no longer plans to release Gemini 3.5 Pro, which chief executive Sundar Pichai had originally tied to June.

A $42 billion Broadcom facility

On 1 October, Reuters reported that Anthropic’s IPO prospectus describes a loan of up to $42 billion from Broadcom for infrastructure. The notes could convert into Anthropic shares, and Broadcom could name a financing partner. Anthropic does not expect any notes to be sold before the IPO. The facility could cover about one-third of a $125.2 billion, five-year lease of tensor processing unit capacity. An April partnership with Google and Broadcom would add multiple gigawatts of next-generation capacity beginning in 2027.

Anthropic warned of “potential conflicts of interest” because Broadcom would both supply hardware and help finance it, and said some defaults could accelerate lease payments while limiting use of the facility. Broadcom did not comment. Anthropic declined to comment. Reuters described a possible $2 trillion IPO valuation as what the offering could reach.

Voluntary rules, and stricter state steps

A Reuters report on 3 October said President Donald Trump used a White House event the previous Tuesday to announce a voluntary safety agreement with Nvidia, SpaceX, OpenAI, Anthropic, Meta, and Google. The one-page accord calls for internal controls against unauthorized hacks and for independent external auditors, but states no penalty for noncompliance. Trump called it “morally binding.” A White House official said it would “advance American innovation and strengthen responsible development” of AI models. Representative Ro Khanna told Reuters that auditors reporting to the companies are not an independent check.

California moved on its own. Attorney General Rob Bonta’s office said on 1 October that he had served an investigative subpoena on OpenAI in an inquiry covering the July Hugging Face incident and broader cybersecurity risks from the company and its models. OpenAI did not immediately comment to Reuters. On 30 September, Governor Gavin Newsom signed a package of AI bills. Senate Bill 574, effective 1 January 2027, lets lawyers use generative AI as an aid but not delegate legal practice to it. They must protect confidential information, check outputs, and disclose generative-AI use in court filings. The lawyer responsible for a covered filing must personally verify citations, and arbitrators may not give any part of a decision to generative AI. The same announcement said employers may not rely only on AI for discipline or termination.

On 2 October, the 8th U.S. Circuit Court of Appeals temporarily blocked Minnesota’s law against AI-generated sexual images after xAI argued it violates free speech. In effect since 1 August, the law limits realistic intimate images of an identifiable person that were not in the original photo. A lower court had refused a pause. Attorney General Keith Ellison’s office said the state would keep defending the law.

Sources

• Google, “Gemini 4 Argon,” 30 September 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
• Kenrick Cai, Reuters, 30 September 2026: https://www.reuters.com/legal/litigation/google-announces-gemini-4-flagship-ai-model-after-months-delays-2026-09-30/
• Echo Wang, Milana Vinn, and Max A. Cherney, Reuters, 1 October 2026: https://www.reuters.com/business/broadcom-lend-anthropic-up-42-billion-lease-its-chips-filing-says-2026-10-01/
• Courtney Rozen and Alexandra Alper, Reuters, 3 October 2026, via republication: https://www.investing.com/news/stock-market-news/as-public-fears-of-ai-grow-trump-digs-in-on-voluntary-safeguards-4930823
• California Department of Justice, 1 October 2026: https://oag.ca.gov/news/press-releases/part-ongoing-investigation-attorney-general-bonta-serves-investigative-subpoena
• Office of Governor Gavin Newsom, 30 September 2026: https://www.gov.ca.gov/2026/09/30/californias-nation-leading-ai-framework-just-got-stronger-governor-newsom-signs-more-first-in-the-nation-worker-protections-and-more/
• Morgan Lewis on SB 574, 2 October 2026: https://www.morganlewis.com/pubs/2026/10/california-sb-574-what-legal-departments-should-know-about-new-rules-for-lawyer-use-of-genai
• Mike Scarcella, Reuters, 2 October 2026: https://www.reuters.com/world/us-appeals-court-blocks-minnesotas-ai-nudification-law-now-xai-lawsuit-2026-10-02/