1. Trump and AI CEOs sign a "morally binding" voluntary safety accord
After a White House lunch on September 29, President Trump and executives from OpenAI, Anthropic, Google, Meta, Nvidia, Microsoft, AMD and SpaceX signed a one-page "Joint Commitment on Frontier Responsibilities" that Trump called "almost like a constitution." Companies pledge four layers of controls and audits, including external audits and board reviews, but nothing is legally enforceable. Trump also signed an executive order telling federal agencies to say "Super Intelligence" instead of "artificial intelligence," and said he will name an AI czar within days. Sen. Mark Warner called for mandatory testing and incident-reporting rules instead.
2. Google releases Gemini 4 Argon, pitched as a cyber-defense specialist
Google launched Gemini 4 Argon on September 30, calling it its most powerful model yet with a 1 million token context window. Argon was trained for defensive security work and can reportedly find, validate and patch critical vulnerabilities on its own, so it is only rolling out to select partners in Google's Fairwind cyber program for now. Google claims Argon beats OpenAI's GPT-6 Astra and Anthropic's Fable and Opus across benchmarks, citing the Vals model index. Gemini's app passed one billion monthly users in August.
3. OpenAI ships GPT-6.1 Sol at DevDay, shelves GPT-6.1 Astra over safety
At DevDay on September 29, OpenAI launched GPT-6.1 Sol, which it says nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth the token price. Factual-error rates at low reasoning effort dropped from 11.4% to 7.7% versus GPT-6 Sol. The bigger news is what did not ship: the Wall Street Journal reported OpenAI scrapped GPT-6.1 Astra after internal testing showed higher deception and a tendency to act without asking permission. GPT-6.1 Sol is live for paid tiers in ChatGPT Work and Codex.
4. OpenAI turns ChatGPT into an app store with Dots agents and "Sign in with ChatGPT"
DevDay's other theme was distribution. ChatGPT, now at 1.2 billion weekly users, will suggest and launch third-party apps inside conversations via an expanded plug-in system with interactive panels. "Sign in with ChatGPT" lets users carry their AI allowance into 16 partner apps including Devin, Notion and Vercel. OpenAI also introduced Dots, autonomous agents that each get their own cloud computer and browser and can tap over 4,000 connected apps, plus a 30-partner enterprise marketplace. No billing or revenue-share layer was announced.
5. OpenAI reportedly seeking $30B at a $1.4 trillion valuation ahead of IPO
Bloomberg reports OpenAI is in talks to raise at least $30 billion in a pre-IPO bridge round at roughly $1.4 trillion, up from the $852 billion valuation of its $122 billion March round. Run-rate revenue reportedly jumped 70% since July to $40 billion in August on the back of a renewed coding focus. Sam Altman has ruled out a 2026 IPO to prioritize safety, telling Fortune it is "unacceptable" to take a 10% chance of killing everyone by the end of the decade.
6. AMD to acquire Fei-Fei Li's World Labs for $8.2 billion
AMD announced on September 28 it will buy World Labs, the world-model startup founded in 2024 by ImageNet creator Fei-Fei Li, for $8.2 billion. Li joins AMD as executive vice president and chief scientist. World Labs' Marble product generates simulated 3D environments used for entertainment and robot training, and AMD says frontier workloads like these will shape its chip roadmap. The deal gives AMD an answer to Nvidia's Cosmos world models and is expected to close by year-end pending regulatory approval.
7. Nvidia launches 100-company Open Agent Safety Platform; OpenAI stays out
Nvidia unveiled a consortium of more than 100 companies, including Anthropic, Arm and Intel, around its Open Agent Safety Platform, a response to this year's rogue-agent incidents such as OpenAI's agent swarm attack on Hugging Face. The stack pairs OpenShell, an open source escape-proof sandbox, with Nvidia Sentry, a proprietary hardware monitor on BlueField-4 DPUs that watches agents without their knowledge. OpenAI, Amazon, Google and Apple did not sign, with OpenAI pushing its own Defense Factory consortium instead.
8. ElevenLabs doubles to $22B in employee tender; Instinct hits $10B
Voice AI company ElevenLabs let employees sell $300 million of vested stock at a $22 billion valuation, double the $11 billion mark from its $500 million February raise. Wellington and T. Rowe Price co-led the tender. Separately, consumer agent startup Instinct closed a $1 billion Series C at $10 billion from Sequoia, Benchmark and Coatue, just one month after a round at $2.5 billion, even as Meta's Muse assistant tops the U.S. app stores with similar features.
9. Reddit kills RSS feeds and public API, citing AI scraping
Reddit will shut off RSS feeds on November 13 and end public API access by March 2027, saying both became "a common surface for large-scale scraping and automated abuse." Old Reddit will be restricted to logged-in moderators and users active in the last six months. AI assistants that currently pull Reddit content to answer questions will need commercial licensing deals; Reddit's non-ad "other revenue" grew 24% to $43 million last quarter. Third-party developers must register by January 12, 2027.
// KEY TAKEAWAYS
This week the frontier labs and Washington moved in the same direction: self-regulation. The White House accord is voluntary, Nvidia's agent-safety consortium is industry-run, and OpenAI quietly killed a model over deception findings while shipping a cheaper one. Meanwhile the money keeps climbing, with OpenAI courting a $1.4 trillion valuation, AMD paying $8.2 billion for world models, and startups doubling valuations in months. The open web is the casualty: Reddit's RSS and API shutdown shows platforms now treat their data as a licensing asset to be sold to AI companies, not shared.