Week in Review: Gemini 4 Argon, Giant Open Models and AI in the Linux Kernel

It was a week where the models got bigger, the lawsuits got more numerous and Linus Torvalds tried an AI coding tool. Here’s what we covered from October 13 to 17, what we’ve learned since, and what we think mattered most.

The big one: Gemini 4 Argon

Google opened the week by shipping Gemini 4 Argon everywhere at once: the Gemini API, enterprise tools, its Antigravity coding environment and the Gemini app for top-tier subscribers. It comes with a 1M-token context window and an introductory price of $2 per million input tokens and $10 per million output that doubles after November 30.

The independent numbers so far are good, with caveats. Artificial Analysis scored Argon 53 on its Intelligence Index, tied with OpenAI’s GPT-6 Astra and behind Anthropic’s Claude Opus 5.5 at 58. It also found Argon uses more than twice as many output tokens per task as Astra, which eats into the price advantage.

Our read: a serious frontier model with unusually specific receipts, a teaser rate you shouldn’t budget around, and a headline benchmark table run at maxed-out settings.

Models: bigger, more open, and some retiring

• Open-weights week. Meta, Reflection and DeepSeek dropped huge models within three days. They include Meta’s 816B Muse Spark Open, which runs on two Mac Studios, Reflection’s 501B Apache-licensed Beam, and an MIT-licensed DeepSeek refresh. Independent scores put DeepSeek on top and at a fraction of the price. Read the licenses before you get excited.
• Microsoft’s hedge. Microsoft AI finished pre-training MAI-Base-2, a 2.8 trillion-parameter mixture-of-experts model trained on 16,000 GB200s in 38 days. CEO Mustafa Suleyman told staff it beats the base model behind OpenAI’s GPT-5.6 Sol on a quarter of the compute. Nobody outside Microsoft can check that yet, and there’s no release date or price.
• Sunset for GPT-4o. OpenAI gave its GPT-4o and GPT-4.1 snapshots three months to live in its API. They leave January 14, 2027, with gpt-6-luna as the suggested replacement. That’s half the six-month minimum notice OpenAI’s own policy promises, and fine-tuned models built on them die the same day.

AI writes kernel code, carefully

Our favorite story of the week: Linus Torvalds posted his first LLM-assisted patch series, a page-cache change made with Claude Code that hit 23x read throughput in a synthetic test. Kernel reviewers then found a race condition before anything merged.

Since then, Meta’s Kiryl Shutsemau posted an 18-patch rework of Torvalds’ series that addresses the review problems. The core patch keeps Torvalds’ authorship and an Assisted-by: LLM tag.

That’s the system working, and a fair summary of where AI coding tools stand: fast at getting a hard idea into code, and still dependent on humans to catch what it got wrong.

Agents meet resistance

In China, 151 apps, including WeChat, Taobao and the game Honor of Kings, opted out of ByteDance’s Doubao Phone Assistant, the agent that reads your screen and taps through apps on your behalf. It doesn’t technically need their permission. ByteDance asked anyway, after a December 2025 standoff in which apps locked out phones running it.

Platforms don’t love being reduced to a backend for someone else’s AI. The fights are about liability for agent-made payments and about who owns the customer. Expect the same negotiation everywhere agents try to use apps the way people do.

Hardware and robots

• TSMC posted $46.7 billion in Q3 revenue, up 41% year over year, with a 66% gross margin and its sixth straight quarter of record profit. It guided Q4 to $52.3 billion to $54.3 billion. The AI boom’s toll booth keeps collecting, whoever wins the model race.
• Nvidia’s RTX Spark laptops went on sale with up to 128GB of unified memory, starting at $2,599. Early reviews found a strong GPU, a CPU that trails Apple’s, battery life under a full workday and local AI that still feels like a developer beta. The 128GB configurations cost roughly $5,900 to $7,000.
• Waymo got Ontario’s go-ahead to start testing in Toronto, its first Canadian city, aiming for public rides in 2028. The province says Waymo needs a credible jobs plan before it moves past human-driven mapping. Winter will be an education.
• Tesla finally showed the second-gen Roadster, nearly nine years after the first tease, with SpaceX cold-gas thrusters, a 2028 production promise, deposits of up to $250,000 and no price. The stock rose premarket the next day, then reversed and was down about 3% by early afternoon. We’ll believe the date when we see the VIN.

Policy

Eight states sued OpenAI over ChatGPT’s alleged harms to children, each in its own court under its own consumer-protection law. Arizona’s complaint calls the chatbot “uniquely sycophantic” by design. Kansas AG Kris Kobach summed up the mood: “We are not going to hold our breath” for Congress.

OpenAI points to its parental controls and a teen mode it launched last month. The states’ bet is that eight courts and eight discovery processes put more pressure on the company than one federal case.

What the week had in common

Three threads ran through almost every story.

• Self-reported numbers. Google, Meta, DeepSeek, Microsoft and Nvidia all led with their own benchmarks. Where independent tests arrived, they mostly confirmed the direction and trimmed the claims.
• Who gets to say no. Kernel reviewers, Chinese app platforms, Ontario’s transport ministry and eight attorneys general all acted as brakes this week. None of them is a federal AI regulator.
• The cost of “frontier.” Argon’s token appetite, terabyte-class open models and $7,000 laptops all point the same way. Capability is getting cheaper per unit and more expensive to actually use.

Our read

The theme of the week was capability racing ahead of control. Models are getting bigger and more autonomous. The checks are coming from kernel reviewers, app platforms and state attorneys general rather than any coherent rulebook.

The takeaway

Watch the gap between what AI can do and who gets to say no. Next week, we’re watching AI security incidents and Tesla’s third-quarter earnings on October 21, and checking whether Gemini 4 Argon’s launch claims hold up under more independent testing.

Image: Dietmar Rabich / Wikimedia Commons, CC BY-SA 4.0