Google's $2 frontier model, a $1 trillion chip rout, and a rulebook for AI designers

Google launched Gemini 4 Argon on 13 October, its new top model for reasoning, coding and long agent tasks. The announcement, written by Koray Kavukcuoglu, Google's chief AI architect, says Argon is available through the paid tier of the Gemini API in Google AI Studio, in the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers, in the Antigravity coding tool for Ultra subscribers, and through Gemini for business. Google says the model answers more directly, hedges less, and reports its progress more often on long tasks.

The headline number is price. Google is charging an introductory $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After 30 November that rises to $4 and $20, the same list price Meta set for Muse Wonder this week. Google cites a 1-million-token limit shared between input and output, which means very long documents or very long answers in a single request. Independent tester Artificial Analysis lists a 1-million-token context, output of about 100 tokens per second in AI Studio, and roughly 42 to 45 seconds before the first answer token, because the model reasons before replying.

Google's own benchmark claims are strong but should be read as vendor-reported. It says Argon ranks first on the Vals Index, scores 87% on the ERQA embodied-reasoning test, and cut indirect prompt-injection attack success to 0.7% on Gray Swan's benchmark. Inside Google, it says agentic code submissions in Antigravity rose more than 35% while code rollbacks fell by nearly a third. Google is also offering 10,000 scientists six months of free top-tier Antigravity access, and giving vetted cybersecurity defenders access with fewer guardrails through a programme called Fairwind.

There are caveats. Google names no competitors in its comparisons, and most of the figures have not yet been reproduced by independent testers. The free API tier is not included, so developers have to pay from the first token. The low price is explicitly temporary, which is why Forbes framed the launch as putting a deadline on Google's price advantage. And per-token price is not cost per task: reasoning models can produce long hidden chains of thought, so a cheaper token can still mean a bigger bill if the model writes more of them.

For businesses, the practical point is that frontier-level capability now costs a few dollars per million tokens, and the next six weeks are the cheapest time to find out whether Argon beats the model you already use. The 1-million-token limit makes whole contracts, tender packs and code bases fit in one request. The 40-second wait before answers means Argon suits background and batch work better than live chat. Teams that measure cost and quality on their own tasks, rather than trusting launch charts, will make better decisions when the price doubles.

Why it mattersA frontier model at $2 per million input tokens lowers the cost of building AI products, but the built-in price rise means anyone who switches needs to budget for the $4/$20 rate, not the launch price.

AI stocks had one of their worst weeks of the year. CNBC reported on Thursday 15 October that 20 of the most valuable AI and semiconductor stocks had lost more than $1 trillion in market value so far this week. Asia took the heaviest hit on Thursday. South Korea's Kospi index closed 7.57% lower at 5,554.23 after falling as much as 9% during the day, which triggered a 20-minute trading halt. Japan's Nikkei 225 fell about 4.9%, and Taiwan's main index dropped more than 3%.

The memory-chip makers that supply AI data centres fell hardest. Closing figures reported by TradingKey and cited by Gotrade put SK Hynix down 11.06%, Samsung Electronics down 9.48% and SoftBank Group down 9.33%. Bloomberg, as reported by Gulf News, said SK Hynix is now 57% below its summer peak after rising more than 1,000% in the 12 months to June. A US chip-stock gauge lost about $1.6 trillion over six trading sessions, and foreign investors sold more than $10 billion of Korean shares in October.

The trigger was the cost of money, not a collapse in AI demand. US consumer prices rose 3.7% in the year to September, up from 3.4% in August, reviving expectations that the Federal Reserve could raise interest rates. The US 10-year Treasury yield traded around 5.1% on Thursday after reaching 5.35% earlier this month, its highest since 2002. AI data centres are built with huge amounts of borrowed money, so higher rates make every new project more expensive. Bloomberg pointed to strain in CoreWeave's bonds, with its 2032 notes near 78 cents on the dollar, and to record yields SoftBank paid to fund its AI plans.

It is important not to over-read one bad week. TSMC, which makes most of the world's advanced AI chips, reported better-than-expected profit and called AI demand very strong, even as its shares slipped about 2%. Lorraine Tan of Morningstar told the AP that any risk to perceived AI spending growth could slow chip demand in markets such as Korea and Taiwan. That is a risk, not a forecast. Some of the fall also reflects how far these stocks had run up, and heavy leveraged retail trading in SK Hynix.

For ordinary businesses, the message is about costs. If financing data centres gets more expensive, AI providers have less room to keep prices low and more pressure to show profits. That could mean fewer free tiers, more introductory prices that later rise, as with Google's Argon this week, and tighter limits on heavy users. Companies that know what they spend on AI, which tasks actually pay back, and how easily they could move to a cheaper model will be in a stronger position than those running on habit.

Why it mattersWhen borrowing gets expensive, the companies building AI infrastructure face pressure to raise prices or cut free usage, which flows through to every business that relies on AI tools.

Figma used its first Config event in India, held in Bengaluru this week, to announce Specs, a new way to tell AI agents how a company's products should look and behave. Specs is a Markdown-based editor. A team writes down its design language, such as colours, spacing, tone of voice, component rules and do's and don'ts, in a plain-text document. Both designers and AI agents can then refer to it. The document can be its own file or sit directly on the Figma canvas next to the designs it describes.

According to TechCrunch, a spec can link to elements on the canvas and embed FigJam diagrams, code layers and designs. Teams can also combine design libraries, code libraries and written guidelines into one central source of design-system context, which Figma's own agent and outside agents consult before they act. Figma also moved Code Layers, first shown in private beta in June, into public beta. It lets users ask the agent to turn a design into working code for an interactive prototype and then adjust parts of that code.

Figma is also widening the set of AI tools that can work with it. Its MCP connections, the standard way AI agents plug into other software, now support more than 20 new tools, including Open Code, Dia, Grok Bot and the Indian AI startup Sarvam. Payload, the open-source content management system Figma bought last year, is being built into Figma Make so users can generate dynamic websites and manage their content in one place. Figma says India is its largest market by users outside the US.

Figma is careful about what Specs promises. Because AI output is unpredictable, the aim is to give agents more context so their results are more consistent, not to guarantee they follow every rule. Coverage of the launch did not include pricing or a general-availability date for Specs, and Code Layers is still a beta, so features may change. Loredana Crisan, Figma's chief design officer, told TechCrunch that 17 times more lines of code are produced today than before, but only 1.3 times more software releases.

The bigger shift is that design rules are becoming instructions for machines, not just reference pages for humans. As more screens and web pages are generated by agents, the quality of the written rulebook decides whether the output looks on-brand or generic. That creates demand for people who can turn a brand's visual identity into clear, testable rules, and for teams that can connect design files, code and agents into one reliable workflow. Markdown also travels well: the same file can guide coding agents outside Figma.

Why it mattersAs AI agents start producing real screens and websites, a clear written rulebook becomes the difference between on-brand work and generic output, and somebody has to write and maintain it.