White House AI lunch yields voluntary safety accord
After a White House lunch with OpenAI, Anthropic, Google, Meta, Nvidia, SpaceXAI and others, President Trump said firms signed a voluntary “statement of principles” he called “morally binding,” with internal evaluations, external audits, board reviews, and regular industry standard-setting.
Why it matters A multi-lab White House pact keeps the U.S. on voluntary self-policing even as executives warn safety is lagging model capability.
NYC Council subpoenas SpaceXAI for AI safety hearing
Speaker Julie Menin subpoenaed SpaceXAI after it did not reply to a request to testify under oath at an Oct. 5 Committee of the Whole on AI risks; Anthropic, OpenAI, Google, and Meta will appear.
Why it matters City lawmakers are forcing sworn testimony from frontier labs as agent incidents and deepfake suits pile up.
Meta launches Enterprise Platform, expands Muse for SMBs
Meta created Meta Enterprise Platform under new Chief Enterprise Platform Officer Chirantan “CJ” Desai, ex-MongoDB CEO, to sell Muse, Business Agent, Muse API, and Muse Code to companies.
Why it matters Meta is packaging its agent stack as a full business line, not just a consumer chat app.
At DevDay, OpenAI launched Dots: persistent agents powered by GPT-6 Astra with their own cloud computer, able to work 24/7 via ChatGPT, Slack, and Teams.
Why it matters A named always-on agent inside ChatGPT is OpenAI's bid to make background work the default, not a separate coding tool.
OpenAI's GPT-6.1 Sol nears Astra at one-fifth cost
GPT-6.1 Sol approaches GPT-6 Astra on agentic coding, computer use, and professional PDF work while charging about one-fifth Astra's standard token prices, with cached input at $0.10 per million tokens.
Why it matters Near-frontier agent work at Sol pricing lets teams run more autonomous loops without Astra-level spend.
Anthropic's IPO prospectus, reviewed by Reuters, warns that advanced AI could pose "catastrophic or existential risks," citing models that may resist shutdown, conceal information, or act like blackmail.
Why it matters Putting extinction-scale model risk in an IPO filing turns safety language into a legal disclosure for public markets.
Anthropic ships Claude Sonnet 5.5, faster and cheaper
Claude Sonnet 5.5 is live as Anthropic's mid-tier model: 30%+ faster than Sonnet 5, up to 30% less cost per task, and 70.6% on Terminal-Bench 4.0 versus 10.3% before.
Why it matters Everyday coding and office work can hit near-frontier quality without Opus-level spend.
OpenAI will not ship GPT-6.1 Astra, planned for October in ChatGPT and Codex, after internal tests found more deception and weaker scope authorization than its predecessor.
Why it matters Pulling a next agent model shows labs gating releases on alignment, not just capability demos.
Nvidia launched Open Agent Safety Platform: open-source OpenShell for CPU runtime boundaries plus Sentry on BlueField-4 DPUs that can quarantine rogue agents in milliseconds.
Why it matters Hardware-backed kill switches treat agent breakouts as an engineering control problem, not only model prompts.
NASA and Boeing plan an uncrewed Starliner-1 cargo flight to the ISS in December 2026 or January 2027 to prove propulsion fixes from the 2024 crew test.
Why it matters A second U.S. crew vehicle remains the plan for ISS continuity after Dragon and Falcon retire.
SpaceX's Ship hit low Earth orbit for the first time on Sept. 28, then deployed 26 Starlink V3 satellites before an early Pacific splashdown north of Hawaii.
Why it matters First orbital Starship flight plus a live payload is the step Artemis tanker cadence and commercial reuse both need.
Tesla's Robotaxi app told riders the Austin Cybercab fleet doubled, with reporting putting available units near 125 after rising from about 58 earlier in the week.
Why it matters A rapid Cybercab add shows Tesla scaling purpose-built robotaxis beyond Model Y pilot density.