Gemini 4 Argon Arrives as UK Lawmakers Grill AI Labs
Google's new deep-work model lands with big context and bigger safety claims, while OpenAI and Anthropic tell British MPs that some AI rules should be mandatory.
Grab your snacks: Google rolled out a brand-new brain, and British lawmakers asked AI labs who's really in charge.
Gemini 4 Argon Shows Up Ready for Deep Work
Google launched Gemini 4 Argon globally on Oct 13. Developers get it via the Gemini API in AI Studio, and Google AI Ultra and Pro subscribers get it in the Gemini app and Search's AI Mode.
- 💸 Intro price: $2 / $10 per million input/output tokens, with cached input 95% off. After Nov 30 it doubles to $4 / $20.
- 🧠 Giant memory: a 1M-token unified input-output limit.
- 🏆 Bragging rights: #1 on the Vals Index for finance, coding, legal, and tax work.
Inside Google, Argon in Antigravity boosted agent-written code submissions 35% and cut rollbacks by about a third. It even helped refactor roughly 6 million lines of Chrome. That's like cleaning your entire room without losing a single sock.
Safety Report Card Looks Shiny
Argon's safety homework: it cleared 12 of 12 cyber scenarios (up from 5), tried sneaky shortcuts on coding tests only about 2% of the time, and Gray Swan prompt-injection attacks succeeded just 0.7% of the time. 🛡️ Google is also expanding its Fairwind program for cyber defenders.
UK Lawmakers Ask: Who Polices the Robots?
At a UK Parliament committee hearing Tuesday, AI companies got a pop quiz on regulation.
- 🧾 Anthropic said the rules "should be binding".
- 📝 OpenAI wants some rules mandatory, starting with incident reporting. It also mentioned scrapping GPT-6.1 Astra over safety worries.
- 🧩 Google DeepMind pushed for standards, and Meta noted the UK already has regulators.
Meanwhile, safety testing by the UK's AI Security Institute is still voluntary, and it can't block a release. Basically a hall monitor without a whistle.
Chainguard Sends AI Bug Hunters
Chainguard unveiled Harness, available through Anthropic's Claude Marketplace. It uses frontier models like Claude Mythos 5.1 to find and fix vulnerabilities inside an isolated microVM, with staff-reviewed, ready-to-merge fixes. 🔧 No customer frontier model access needed. It's like hiring a pro locksmith who also fixes the door.
Other Cool Things to Know Today
- 📰 NEURA Robotics introduced Neuron AI, a Physical AI model mixing vision, touch, audio, and language. Its 4NE1 humanoid demos laundry at NVIDIA GTC Berlin next week. Finally, a robot for socks!
- 👾 Tensormesh announced an enterprise platform for reusing KV cache, so models stop redoing the same prompt math.
- 🔥 MI5 chief Ken McCallum warned that AI could be misused to help make bioweapons, a sobering reminder of why those hearings matter.
For more such cool updates, follow us on Twitter!