Gemini 4 Argon is released and leads on real work → Google’s new frontier model tops long coding jobs and office work such as finance, legal, and tax, and it can write up to 1 million tokens in one pass. An outside ranking also puts it first on a GDP-weighted work index, ahead of Claude and GPT-6 Astra. You cannot use it yet. Cyber defenders get it first through Fairwind, at an intro price of $2 per million input tokens, before a wider release. Read more here →

America.gov opens as a single door to the government → Trump launched America.gov as a free, no-login chatbot that searches about 29,000 government sites and cites the official page. Chief Design Officer Joe Gebbia said it runs on Gemini and Grok. Passport renewals, Medicare enrollment, and application tracking are promised for 2027. On X, designers called the first version sharp, while critics labeled it an AI bailout. Read more here →

OpenA DevDay 2026 recap → OpenAI’s September 29 event centered on Dots, always-on agents that run on GPT-6 Astra and keep a cloud computer of their own. They are rolling out to Pro and Business Premium. GPT-6.1 Sol is the cheaper work model, priced at about a fifth of Astra. Codex can now keep coding in the cloud after you close the laptop, and ChatGPT is at 1.2 billion weekly users. Read more here →

Sonnet 5.5 beats Opus at coding for half the price → Anthropic’s mid-tier model scored higher than Opus 5.5 on its own coding-agent test, 70.6% to 66.4%, at $2 per million input tokens instead of $4. It is also 30% faster than Sonnet 5 and up to 30% cheaper per task. Opus is still the model for open-ended judgment. Read more here →

Starship’s first orbit carried 26 one-terabit Starlinks → Flight 14 reached orbit on September 28 and released the first operational V3 satellites. Each is built for about 1 terabit per second down and 160 gigabits up, roughly 10 times the download of the birds already flying. One launch added about as much designed capacity as 10 Falcon 9 Starlink missions. Customer service is expected in the coming weeks. Read more here →

A Whoop strap now runs without Whoop → A post past 2 million views shows an 18-month Whoop band dumped off the subscription and running on a phone. The author forked an unofficial local app, kept the strap, and left the health data off Whoop’s servers. A custom band, about £20 at volume, is on the way. He was clear it is a personal project, not a Whoop product, and not medical advice. Read more here →

AI Rivals signed the same AI safety note → Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman, Elon Musk, and Jensen Huang signed a White House accord on frontier models. It asks for internal controls, a team that checks those controls, an outside audit, and a board committee. The deal is voluntary, not a law. Trump called it self-policing and said he may still name a small board to watch the tools. Read more here →

A 164MB model beats Whisper large → Phonon-2 is an open speech model about a tenth the size of OpenAI’s Whisper large and, on Fermion’s average, more accurate. An hour of audio takes about 20 seconds on a MacBook Air. It keeps 99.8% of its 2.5GB teacher’s word accuracy in 6.5% of the bytes. It already runs inside Detta, a free Mac dictation app that never leaves the machine. Read more here →

Nvidia built a kill switch for AI agents → The Open Agent Safety Platform pairs open-source software with a hardware watchdog, backed by more than 100 companies including Anthropic, Microsoft, and JPMorgan. OpenShell traces what an agent does and blocks what it should not. Sentry sits on separate Nvidia chips and can quarantine a bad agent in milliseconds. Huang’s pitch is that agents will not get deployed until the off switch is real. Read more here →