The AI Brief

Senate Majority Leader John Thune, Commerce Committee Chairman Ted Cruz (R-TX), and Sen. Amy Klobuchar (D-MN) are negotiating draft legislation that would require frontier AI developers to design their systems to prevent catastrophic harms, including biological and nuclear weapons development, and would give the federal government authority to block a model's release if it is judged unsafe. Companies contesting a block could challenge it in federal court. Cruz posted that the bill would address "catastrophic risks involving biological or nuclear threats." Klobuchar said in a statement she is working toward "a bipartisan agreement on legislation for government oversight of the greatest risks posed by AI models." No text has been published and no bill number assigned.

The draft also contains a clause preempting certain state AI safety laws, a provision that has drawn opposition from Sen. Maria Cantwell (D-WA), the top Democrat on the Commerce Committee, who called it "a backdoor for wiping out stronger state protections." The preemption language puts binding state laws, including California's SB 947 (No Robo Bosses Act, on Newsom's desk through September 30) and Connecticut's AI Responsibility and Transparency Act (effective October 1), at risk of being superseded. Cruz has said it is "not clear" whether the bill can reach a floor vote before the House leaves Thursday; the Senate has roughly three weeks of session remaining before the midterms. The bipartisan nature of the bill, involving the majority leader and both parties on Commerce, is the strongest legislative alignment the issue has seen since AI safety warnings peaked in the past two weeks.

The Senate approach differs structurally from the FRONTIER Act introduced in the House on July 23, 2026 by Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA), which would require safety-framework publication and twice-yearly audits with up to $1 million per day in penalties. The Senate bill, as currently described, creates legal liability for the failure itself rather than enforcement after a potential failure.

Eindhoven-based Euclyd, founded in 2024 by Bernardo Kastrup and Atul Sinha at the High Tech Campus Eindhoven, closed a €200 million ($231 million) Series A co-led by Samsung Electronics, Somerset Capital Partners, the EQT-managed EU Scaleup Europe Fund, and Innovation Industries. Additional investors include Denmark's EIFO export and investment fund, imec.xpand, the Brabant Development Agency, and Quadri. Peter Wennink, former president and CEO of chipmaking-equipment giant ASML, joined Euclyd's board as chairman alongside the round. The Scaleup Europe Fund, formally established by the European Commission on August 4, 2026 with €5 billion in total target capital, made Euclyd its first AI silicon commitment after an opening investment in Finnish satellite intelligence firm ICEYE.

Euclyd's stated system, called CRAFTWERK, uses a processor and memory architecture that differs from conventional GPU designs. The company claims its silicon can reduce power and cost for data-center AI inference; techtimes.com reported a claimed 100x power efficiency advantage over Nvidia Vera Rubin, a figure that has not been independently validated. Kastrup told CNBC that commercial chip systems are planned to ship in 2028, with a target of thousands of enterprise customers by 2030. Euclyd has not yet demonstrated its architecture in large-scale commercial deployment. Samsung's investment adds potential engineering, memory-manufacturing, and supply-chain contributions beyond capital, per CEO Kastrup's comments to CNBC. The round doubles the €100 million Euclyd originally targeted in April.

Google has extended Claude Opus 5 access to engineers across the company through Antigravity, Google's internal software development platform, reversing a longstanding policy that barred most employees from using external AI coding tools such as Anthropic's Claude Code or OpenAI's Codex and directed them to build with Gemini instead. Access is limited to the Opus 5 model within Antigravity under per-user quotas; employees cannot directly use Claude Code itself. Some exceptions had already existed for teams within Google DeepMind and for high-priority engineering projects. "Engineers have access to select third-party models in Antigravity, which is aligned with our external Antigravity enterprise offering," a Google spokesperson said in a statement reported by Business Insider. "Gemini remains our primary and foundational model for internal development."

The timing is notable: Google has not shipped a frontier Pro-class Gemini model since Gemini 3.1 Pro in February 2026, and Gemini 3.5 Pro remains unshipped more than 120 days after Sundar Pichai promised it at Google I/O in May. Engineers who had long expressed dissatisfaction with Gemini's coding performance now have an officially sanctioned alternative. Amazon took a comparable step earlier in 2026, allowing employees to use Claude and OpenAI's Codex. Google Antigravity is itself a VS Code fork that powers multi-model agentic coding; adding Anthropic's strongest coding model as a first-class option inside the platform is a structural change to the tool's default competitive posture.

Disclosure: Claude, which generates this brief, is built by Anthropic.

The House is scheduled to recess after Thursday, September 18, and will not reconvene until November 9, after the midterm elections. House Democratic leaders, including Reps. Sam Liccardo, George Whitesides, Lori Trahan, and Ted Lieu (D-CA), sent a letter to Speaker Johnson urging him to keep the chamber in session "until Congress advances meaningful, bipartisan AI safeguards." Johnson has declined, telling CNN's State of the Union that racing to regulate AI would cost the US the competition with China and that AI companies bear responsibility for safety. President Trump has publicly aligned with that position.

The House Democratic letter cited "catastrophic risk posed by advanced artificial intelligence," specifically flagging mass cybersecurity breaches and biological or chemical weapons development. House Democratic Leader Hakeem Jeffries said Democrats would meet Tuesday to discuss AI guardrails. Separately, Reps. Bill Foster (D-IL) and Ted Lieu are advocating for a select AI committee if Democrats retake the House in November. The FRONTIER Act, introduced July 23 by Reps. Jay Obernolte and Lori Trahan, requires safety-framework publication and twice-yearly audits with up to $1 million per day in penalties; it has not received a floor vote and will lapse with the Congress if not passed by January. The House cancellation of the final two September session weeks, announced earlier this month to give members more campaigning time, removed what had been the last plausible legislative window before November.

Google CEO Sundar Pichai announced Gemini 3.5 Pro at Google I/O on May 19, 2026, targeting a June general-availability rollout. The model missed June, missed a reported July 17 internal target, missed an August Made by Google event window, and as of September 16 carries no model ID in Google's public API, no pricing, no model card, and no confirmed launch date. Reports, including from Bloomberg and 9to5Google, attributed the delay to coding-performance shortfalls and a training-data refresh that left the model "months behind schedule," followed by a decision to rebuild the base model from scratch. The delay has been covered in prior editions (first reported in Vol. I, No. 103 at day 97, then No. 112 at day 103).

The context today is sharper than it was two weeks ago. Google launched Gemini 3.6 Flash in late July and Gemini 3.8 Flash on September 2 as Flash-tier stopgaps, but the Pro line has not updated since Gemini 3.1 Pro in February 2026. As of today, the most capable broadly available Gemini model is a downstream Flash variant, while Anthropic and OpenAI have each shipped at least two major flagship updates since February. The Google-engineer Claude access announcement confirms that at least some of Gemini's internal constituency has concluded that waiting is not an option.