AI news, 16 September: OpenAI, Anthropic and Google coordinate on AI safety

OpenAI, Anthropic and Google deepen safety talks. OpenAI’s global policy chief Chris Lehane confirmed that the company has been working with rivals Anthropic and Google DeepMind on AI safety measures for several weeks, according to reports from Bloomberg, Reuters and TechCrunch. The discussions, which predate Anthropic CEO Dario Amodei’s recent public call to pace frontier development, aim at shared standards and risk mitigation. Lehane said the firms do not believe they need an antitrust waiver to collaborate on safety and that OpenAI supports bipartisan efforts for independent verification of advanced models. The move comes amid intense public debate over catastrophic risks, with industry leaders divided and political figures pushing back against slowdowns.

Google launches Gemini 3.8 Live voice models. Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its most advanced native speech-to-speech models yet, available via the Gemini API, AI Studio, Search Live and select Workspace apps. The Extended Thinking version tops Artificial Analysis’ Speech-to-Speech Quality Index and leads agentic voice benchmarks while supporting background tool use, visual grounding and fluid multilingual conversation. Priced competitively for developers, the models enable more natural real-time voice agents that reason and speak simultaneously, according to Google’s announcement.

Salesforce and Nvidia debut Koa enterprise model. At Dreamforce, Salesforce and Nvidia announced Koa, a specialized CRM reasoning model built by post-training Nvidia’s Nemotron on synthetic data reflecting nearly three decades of Salesforce workflows. Designed for multi-step agent tasks in sales, service and support, Koa runs entirely within Salesforce’s infrastructure with no customer data used in training. Early benchmarks show it matching or beating general frontier models on CRM actions with far fewer errors; it is in pilot now with general availability expected later in 2026.

Meta rolls out Meta One AI subscriptions. Meta launched Meta One, a tiered subscription service offering expanded AI usage—such as more Muse image and video generation and Instagram Restyle edits—plus premium features across Facebook, Instagram and WhatsApp. Individual bundles start around $8–20 monthly, with higher creator and business tiers; the core free experience and basic Meta AI remain available. Meta said the plans, already tested with millions of trials, help diversify revenue while keeping everyday access open.

Microsoft publishes Humanist AI Code of Conduct. Microsoft AI released a draft “Humanist AI Code of Conduct” for public consultation, emphasizing that people matter more than AI and that models must remain under human control. Absolute constraints bar assistance with cyberattacks, weapons, deepfakes or deceptive evasion of oversight. The document will guide future MAI model training after feedback; current models are not yet trained on it. The release responds to growing safety concerns across the industry.

Other developments. OpenAI acquired camera-AI startup Glass Imaging for more than $300 million, according to the Wall Street Journal, adding imaging expertise amid hardware ambitions. The Gates Foundation pledged at least $1 billion over two years to expand equitable AI access in health, education and agriculture. Nvidia CEO Jensen Huang publicly rejected calls for an industry slowdown or new regulations, aligning with recent White House comments while affirming the need for safe engineering.