Good morning. Google’s turn at the frontier: Gemini 4 Argon arrived yesterday and, unlike Meta’s slightly wobbly Muse Wonder launch earlier this week, the early reception is unusually warm — strong benchmarks, aggressive pricing, and some genuinely interesting deployment stories. Also today: an alleged espionage caper at OpenAI involving toilet paper and a USB port, and Character.AI finally files the paperwork on what was effectively a two-year-dead company.
Gemini 4 Argon lands, and people actually like it. Google launched Gemini 4 Argon at $2/$10 per million input/output tokens — roughly a fifth of what GPT-6 Astra and Claude Opus 5.5 charge — with claimed leads across coding, cybersecurity, science, finance, and law. The Verge’s coverage notes availability is currently limited to Pro and Ultra Gemini subscribers. Internal Google deployments reportedly lifted agentic code submissions 35% while cutting rollbacks by a third, and Chrome’s team used it to compress a months-long Chromium refactor into weeks.
The community verdict is more positive than most launches get. HN commenters called it a “clear step change” for reasoning and writing, with one placing it at the Pareto frontier on the arena.ai leaderboard — comparable intelligence to competitors at $0.41/task vs $3.51. German-language and non-Western cultural knowledge both got specific praise, areas where Google has consistently been ahead. The main caveat: image-to-HTML tests put Argon a tier below Opus 5.5 and GPT-6.1 Sol, so it’s not a clean sweep.
Argon for Robotics is more interesting than the demos suggest. Google also released a robotics-tuned variant scoring 87% on the Embodied Reasoning benchmark, with demos of a Spot robot handling trash cans and zero-shot arm manipulation. Reception was lukewarmer here — one HN commenter pointed out Unitree robots are doing parkour while Argon’s demo “has to be hung up by a hook,” and asked why Google’s resources aren’t producing more dramatic results. Hobbyists building their own robots were more enthusiastic, finding the zero-shot visual control approach immediately useful.
Safety story: Argon solved 12 of 12 cyberattack scenarios. The responsible deployment post is worth reading — up from 5 of 12 in the previous generation, which prompted Google to run a limited “Fairwind Program” with trusted cyber defenders before general release. One early use: Wiz used Argon to find a critical bug in hospital software protecting hundreds of thousands of patient records.
An OpenAI engineer, a bug behind a TV, and a chatbot paper trail. The FBI arrested Abdulrahman Alfozan, a 35-year-old Saudi national and OpenAI engineer, for allegedly planting an encrypted audio recorder behind a conference room TV — concealed in toilet paper, plugged into a USB port. Per the SF Standard and Wired, a chatbot account linked to his wife — sharing identifiers with his own — submitted queries about building, obscuring, and remotely accessing recording devices in the week before. The Verge reports a second device was found in a bag Alfozan left behind containing his wallet and ID. FBI Director Kash Patel said he was after communications about “super intelligence technology.”
The subplot everyone noticed. HN commenters flagged two things: Alfozan is reportedly a venture partner at a VC firm alongside the son of a senior Saudi intelligence official, and OpenAI couldn’t resist noting in its statement that “we used our models to understand the device.” One commenter: “Did OpenAI do a marketing bit in the middle of an espionage accusation?” Also worth watching — Alfozan worked on data infrastructure at Meta from 2017 to 2022, and there’s speculation Meta is now auditing that period.
Character.AI files Chapter 11. The platform filed for voluntary Chapter 11 to pursue a court-supervised asset sale, with Disney agreeing to buy some assets for $20 million per TechCrunch. CEO Colin Adams warned the platform may wind down if no broader bid materializes. The company has effectively been a shell since August 2024, when Google paid $2.7B to license its tech and hire Noam Shazeer and Daniel De Freitas back. One HN commenter calculated the leftover entity was burning ~$1M/month on 86 Bay Area employees averaging $278k base — plus nearly $100k owed to DoorDash.
The data question nobody is answering. Several commenters raised the obvious concern: what happens to the deeply personal conversations millions of users had with their AI “friends”? Character.AI’s user data is arguably more intimate than most social platforms, and bankruptcy proceedings have historically been where privacy policies go to die. Worth watching how the asset sale handles it.
That’s the morning. Enjoy watching the Argon benchmarks get picked apart in real time — and maybe check behind your conference room TV.