AI
Google launches Gemini 4 Carbon, a cybersecurity-focused frontier model
Two weeks after Argon, Google's new model tops the Artificial Analysis index at 59, but most people cannot use it yet and its launch pricing is temporary.
Still
Crunchy.
YESTERDAY, STILL WORTH KNOWING
Google launches Gemini 4 Carbon, SK Hynix predicts memory shortages until 2030, and Epic confirms AI now writes much of its engine code.
Google doubles up on frontier models
Google has launched Gemini 4 Carbon, a cybersecurity-focused frontier model released just two weeks after Gemini 4 Argon. The company says Carbon tops version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, up from Argon's 50, and reports leading cybersecurity benchmarks for finding vulnerabilities, writing exploits and producing patches. Google also says Carbon blocked every attempted prompt-injection attack in one evaluation, though The Decoder observes that test appears to be saturated. Independent results show trade-offs: on Vals AI Carbon places second behind Claude Sonnet 5.5, its hallucination rate rose from 4% to 11%, and some agentic tasks take far longer. Access starts this week with a subset of cyber defenders, with wider availability through Google's Fairwind Program, and pricing matches Argon's temporary rate of $2 per million input tokens and $10 per million output tokens, set to double once the introductory period ends.
Alphabet's Q3 2026 earnings call gave the launch context: CEO Sundar Pichai reported revenue up 25% year-over-year, Google Cloud up 103%, and the Gemini app reaching one billion users, which he called the fastest-growing product in Google's history. He credited a full-stack approach spanning chips, models and consumer products, and said first-party model APIs now process nearly 32 billion tokens per minute.
The memory squeeze bites
SK Hynix, the Korean memory chipmaker, predicts memory shortages will persist until at least 2030 as AI demand outpaces supply. Quarterly revenue rose 328% year-over-year to about $78.1 billion, with operating profit up 486%, and chief financial officer Kim Woo-hyun said PC and mobile shipments are expected to decrease because of rising memory costs. The company has 14 long-term deals covering more than half its output in 2027 and 2028.
Solidigm, the American flash memory subsidiary of SK hynix, is reportedly offering a $45 billion US plant in exchange for approval to export equipment to upgrade its NAND factory in Dalian, China, Bloomberg reports citing unnamed sources. That ask conflicts with US export controls, and it is unclear how officials received it. The proposal lands in a volatile market where manufacturers have cut consumer NAND production in favour of data centre chips, which was cited as one reason SK hynix cancelled a planned Solidigm IPO this week. The squeeze is also reaching graphics cards: Nvidia's GeForce RTX 5080 now lists at nearly $2,000 across US retailers, roughly double its $999 launch price, driven by a shortage of its 16 GB of GDDR7 memory.
AI in the engine room
Epic Games has confirmed it uses AI tools including Claude Code and Codex to write much of the code for Unreal Engine and Fortnite. Executive VP Marcus Wassmer said engineers are no longer writing the bulk of the code but still guide the tools, and estimated a D3D12 rendering rewrite at three to six weeks took about a week, he says.
Amazon has told its engineers to focus on outcomes rather than maximising AI token consumption, a practice it calls tokenmaxxing, in a memo by VP Terk Terkowitz. It mirrors Microsoft's August guidance, and The Register suggests more enterprises may turn to cheaper open-weight models. Anthropic, meanwhile, launched reporting tools letting enterprise customers see what their AI spending produces, plus pooled group budgets, and reports 900% net revenue retention, meaning existing customers' spending grew ninefold between Q3 2025 and Q3 2026.
Rules, clones and proxies
Valve reportedly told developers at a Korean conference to combat AI-generated game clones on Steam by building communities and adding features, offering no platform crackdown; Eurogamer is seeking clarification and is sceptical of the approach.
Meta may again delay its updated business terms, originally due 30 October then pushed to 20 November, after publishers objected that they appeared to allow AI training on news content. Communications head Andy Stone said implementation would slip if consultation runs past 20 November. New York Governor Kathy Hochul signed a law requiring AI companies to verify users' ages before chatbot use, banning chatbots from posing as real people or encouraging self-harm in minors, with penalties up to $25,000 per violation and effect in six months. In security, ESET researchers report Belarus-linked hackers are increasingly aligning cyber operations with Russia's intelligence priorities in Ukraine, documenting a first destructive data-wiping attack by a Belarus-aligned group in Europe in late May.