Still
Crunchy.

Google returns to the frontier with a price war attached

Google launched Gemini 4 Argon, its first major frontier model since Gemini 3 in late 2025, claiming benchmark wins over GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 across coding, cybersecurity, science, finance and law. The company says it already uses the model internally, from patching security vulnerabilities to halving the compute behind YouTube's comments section. As with recent frontier releases, the public rollout followed a limited test with trusted cyber defenders.

The pricing is the sharpest edge: an introductory $2 per million input tokens and $10 per million output tokens, roughly a fifth of what rivals charge, though rates double in late November and cheaper per-token prices do not guarantee lower real-world bills. Access is also narrowing at the bottom: Argon is available to Pro and Ultra subscribers, while free users were recently restricted to the smallest Flash-Lite model. Meanwhile Shopify opened its billion-listing catalogue to any AI agent, not just partners like Meta's Muse, with buyers approving each purchase, a move that lands in the middle of an ongoing dispute between Amazon and AI firms over whether agents should shop at all.

AI's boundaries debated, from classrooms to airgaps

The UAE became the first country to make AI a compulsory school subject from kindergarten upwards, teaching it as literacy, with young children spotting AI mistakes rather than using chatbots directly. That puts it opposite New York City and Norway, which have restricted student-facing generative AI over wellbeing concerns. A separate open letter signed by 406 scientists, including Geoffrey Hinton and Yoshua Bengio, went further on safety, calling for AI models to be airgapped from the internet, meaning fully disconnected, plus binding international containment standards. OpenAI's Noam Brown countered that even physical separation is not foolproof, since computers could theoretically signal each other through temperature sensors.

Anthropic's red team added a concrete worry: its study found small AI models can set up working environments on cheap hardware, including Android TV boxes costing about $55, but only a large external model completed a full self-replication sequence. It projects a 30-billion-parameter model could close that gap by the end of 2026, though any such worm would be far slower than conventional malware.

Mathematics splits over machine-made proofs

OpenAI's release of 722 AI-generated maths manuscripts, of which only 300 have been formalised into computer-verifiable form, has alarmed researchers who say checking them could take years. Terence Tao noted the results have generated far fewer seminars and collaborations than traditional breakthroughs, while Bryna Kra warned the field could become more secretive. The contrast with China is pointed: a new Beijing alliance led by Fields Medal winner Shing-Tung Yau aims to push AI into research with over 1,000 researchers, while 25 Fields Medal winners in the West have voiced concern about AI proofs. A Johns Hopkins professor also raised the possibility, still a prediction, that such systems could one day threaten encryption.

Hardware and gaming briefs

NVIDIA's RTX Spark laptops go on sale on Friday with its software ecosystem ready, including agent apps, a local Nemotron model and anti-cheat support for online games. Chinese firms have reportedly bought over 1 million Nvidia GPUs to fill a domestic chip shortfall, pushing the RTX 5090 to $7,199 at resellers. Apple's new $299 Apple TV 4K adds hands-free Siri AI and a Find My remote, arriving late after Siri delays. Rockstar is reportedly planning a GTA 6 PC release within four months of the console launch, and Xbox's Asha Sharma has insisted new intellectual property still matters despite layoffs and a franchise-first reset.