AI News Roundup: November Two and the Cost of Autonomy

Overview: Monday delivered three bills at once. The New York Times sued the search index that quietly feeds several AI labs, xAI admitted its always-on agent costs more than it earns, and South Korea put $16 billion of public money behind its own model. Add Anthropic's revised SEC filing, which warns the company may be close to recursive self-improvement, and the day's through line is clear: autonomy now has a price list, and everyone is reading it.

Publishers Sue Brave Over the Index That Feeds AI Agents

The New York Times filed a 72 page complaint in the Southern District of New York against Brave Software, arguing the company runs a crawler nobody names, builds an index of more than 30 billion pages, and sells access to Anthropic, Perplexity, Mistral, xAI and Cohere. Hours later the Daily News, Chicago Tribune and Orlando Sentinel filed a parallel suit in the same courthouse. The claim is not really about a browser. It targets the retrieval layer every agent stack depends on, and asks whether an index that ships paywalled text downstream is a fence for stolen goods.

Anthropic Warns It Is Close to Recursive Self-Improvement

Buried in the revised S-1A that Anthropic filed on Monday, beyond the headline revenue growth and equally steep expenses, is a risk sentence with unusual teeth. The company says its internal estimates and expectations for future models, rather than its current ones, indicate it may imminently meet its Responsible Scaling Policy definition of "dramatically accelerated progress attributable to AI". It also concedes it has in the past failed to notice a significant acceleration in its own development rate until after the fact. No current model has been designated as crossing that threshold, and Anthropic says humans stay in the research loop. Still, this is the first time a frontier lab has written recursive self-improvement into a securities filing as a plausible near-term condition rather than a thought experiment.

xAI Throttles Grok Bot as Always-On Agent Costs Outrun Revenue

Users woke up over the weekend to find their Grok Bot allowances exhausted. Elon Musk's explanation was blunt: serving the agent cost an unsustainably high amount relative to revenue. Grok Bot is not a chat window with a scheduler. It runs inside its own cloud computer, signs into apps the user already pays for, and grinds through multi-step tasks until the job finishes. That architecture is expensive by design, and flat subscriptions do not cover it. This is the first public admission from a major lab that unattended autonomy has a unit-economics problem, and it lands right as every vendor is pushing annual agent plans.

South Korea Puts $16 Billion Behind a Sovereign AI Model

Seoul's 2027 budget, the largest in the country's history, routes 21.3 trillion won, about $15.8 billion, into chips, physical AI and domestic data centers, with another 4.7 trillion won earmarked for a Korean-built frontier model. That AI allocation nearly doubles the 10.8 trillion won committed for 2026, and President Lee Jae Myung paired it with a talent plan in which Samsung Electronics and SK Hynix help design university courses. The strategic detail is the word domestic: this is money for compute the country owns, not for rented capacity from American providers. Sovereign AI stopped being a conference phrase and became a budget line.

42AI Exits Stealth With a Behavioral World Model

Israeli startup 42AI came out of stealth with $62.5 million in seed funding and a model that forecasts what a specific person will do next, including in situations it has never observed them in. It is already live at roughly ten large brokerages, where it is used to anticipate how individual traders react to volatility and to manage liquidity and margin. The funding is the headline; the data pipeline is the lesson. Instead of extrapolating from past actions, 42AI samples off-topic signals such as how a customer behaves through KYC, support calls and money transfers, then converts them into behavioral features. Prediction of human decisions is becoming an infrastructure product with a clean API and a compliance story.

What to Watch Tomorrow

• OpenAI merges ChatGPT's Chat and Work tabs: an email to users confirms the two surfaces collapse into one, and Codex starts drawing from the same token allowance pool per TechCrunch. Watch whether published allowance sizes arrive before December.
• Anthropic's threshold assessment: the S-1A says the company is still assessing the underlying data behind its RSI definition, which means a formal designation, or a refusal to make one, is now a disclosure event.
• Brave's response: the complaint names five AI labs as index customers. Expect either a public statement on crawling policy or silence followed by a motion to dismiss.
• Grok Bot pricing: xAI cut limits without a changelog. If a repriced agent tier appears, it becomes the reference point for the rest of the market.

Builder Impact

• Your retrieval layer is now legal exposure. If a product ships third-party search results to an LLM, understand who crawled the content and under what terms. Self-hosted indexes and licensed feeds stop being optional hygiene.
• Price agent time, not agent seats. Grok Bot proved that a flat fee against an agent that browses, logs in and retries forever is a loss-making contract. Meter execution, cap parallelism, and put a hard budget on every loop.
• Sovereign budgets mean sovereign stacks. Korea's allocation is explicitly for domestic data centers and a Korean model. Expect more localization requirements, more on-prem inference, and more procurement questions about where weights and logs live.
• Behavioral models need data plumbing you probably lack. The interesting part of 42AI is not the architecture, it is the discipline of capturing off-topic signals with consent and turning them into features. Any serious prediction product starts there.
• Self-improvement claims change roadmap risk. If a lab says capability jumps may be noticed after the fact, pin model versions, regression test on a schedule, and never assume next quarter's model behaves like this quarter's.