Gemini 4 Argon Opens to Paying Users and the API Two Weeks After Launch -- at $2/$10 Until November 30, Then $4/$20
Google moved Gemini 4 Argon beyond its cyber-defender preview on October 13, opening it to Google AI Pro and Ultra subscribers, the paid Gemini API, AI Studio and Antigravity. Independent leaderboards put it first on one index and five points behind Claude Opus 5.5 on another.
IMF: $300 Billion of Hyperscaler Equity and $650 Billion of Debt Are Riding on $1.4 Trillion of Future AI Revenue Promises
The IMF's October 2026 Global Financial Stability Report, launched October 13 at the Annual Meetings in Bangkok, names an AI correction as one of three risks that could abruptly tighten financial conditions. It puts numbers on the circular deals between hyperscalers and the AI firms they fund.
Nilekani Calls Doom Predictions for IT Services Firms 'Disaster Marketing' and Says India Should Lead on Scaling AI, Not on Building the Biggest Models
Speaking at the Bloomberg New Economy Forum in New Delhi on October 13, Infosys chairman Nandan Nilekani said demand for IT services firms is rising, not falling. He made the case one day after HCLTech's CEO said AI-led price deflation is only in its early phases.
Janas: A Dependency-Free C Engine Runs an 80B Mixture-of-Experts Model on a 32 GB Laptop by Streaming Experts From the SSD
Janas-LLM drew a Hacker News debate on October 12-13. It keeps MoE weights on disk, reads only the experts each token needs, and caches the popular ones. Its author reports large gains over llama.cpp on models bigger than RAM, measured on a single machine.
Framework Watch: Google ADK 2.12 Stops OAuth2 Secrets Leaking Over Its Run Endpoints, Locks Dev Endpoints to Localhost, and Lets You Filter Tool Output Before the Model Sees It
ADK Python 2.12.0, released October 13, is mostly fixes. The ones that matter most close security gaps in the dev server and run endpoints. It also adds an after_tool_callback for screening tool output, human confirmation on workflow nodes, and a 40% faster first model request in long sessions.
Research Radar: AgentSWE Finds Coding Agents Can't Build the Agent You Asked For -- and Often Say They Did; 'Thinking Inertia' Shows Models Keep Reasoning When Told Not To
Two papers near the top of Hugging Face's daily list on October 13 test whether models do what they are told. AgentSWE checks whether coding agents deliver the agents a requirement describes. Thinking Inertia checks whether a no-reasoning instruction actually stops reasoning.
Trend Watch: Anthropic Replaced Git as Merges Climbed Past 30,000 a Day -- and Says Engineers Very Often Had to Tell Claude 'No'
Anthropic's October 13 engineering post on Scion, an in-house version control system built because agent-driven merge volume outgrew git, drew more than 950 comments on Hacker News. The debate was less about git than about what it means when code gets that cheap.