Tech • AI • Robotics • Game

ENFR

Gemini 4 Krypton leaks, Qwen 4 Max lands, FAA router failure snarls flights

AIMonday, November 16, 2026· 14 videos

Briefing

Audio player

Gemini 4 Krypton leaks ahead

Google appears to be testing a new Gemini 4 Krypton checkpoint in its API and in public model arenas under the alias Gemini 4 Hydrogen. Early users describe a marked step up from Argon and Carbon, especially in web coding, SVG generation, simulation tasks and prompt adherence. Shared outputs include a browser-based CSGO-style game, a 3D rocket engine rendered in a single HTML file and more accurate controller diagrams such as PlayStation 4 and GameCube layouts. If those results hold, Krypton could narrow the gap with Claude Opus 5.5 and higher-end agentic models while competing aggressively on price.

FAA outage spotlights SMARTT timing

The FAA says a router failure during network modernization caused the nationwide ground stop on November 13, which halted departures for 4 hours 21 minutes. More than 3,000 aircraft were already airborne, over 800 flights were diverted, and cancellations and delays ultimately affected thousands more travelers across the US and Europe. Scrutiny intensified because the agency had activated SMARTT across all 22 en route centers just three days earlier, though officials said there was no evidence of sabotage or cyberattack. The disruption has renewed questions about resilience, rollback procedures and oversight of modernization in critical infrastructure.

Qwen 4 Max enters flagship race

Alibaba launched Qwen 4 Max, the first released model in its Qwen 4 family, with multimodal input, tool use, web search and a 1 million-token context window. The company says a new mixture-of-experts design cuts token use by up to 50% and task costs by more than 60% versus Qwen 3.8 Max. API pricing was listed at $1.50 per million input tokens and $4.50 per million output tokens, making it a relatively inexpensive frontier-tier option. Independent early testing, however, suggests it still trails several top rivals on raw performance and overall value despite stronger agentic workflow support.

Anthropic pushes Claude into workflows

Anthropic rolled out a broader set of upgrades for Claude projects and desktop coding tools, aiming to make the assistant more persistent and operational inside ongoing work. New project features include a more proactive coordinator, parallel task handling and better organization around recurring assignments. On desktop, Claude Code now supports prompt-generated custom themes, animated backgrounds and status-reactive visuals in local sessions. The changes reflect a wider industry shift from chatbot interactions toward ambient, project-based assistants that stay embedded in software development and knowledge work.

Claude Code meets Hostinger VPS

A new deployment workflow combines Claude Code, a Hostinger plugin and an MCP connector to move simple websites and apps onto a VPS with minimal manual administration. The setup reduces friction around DNS, SSH, HTTPS, domain setup and other infrastructure tasks that often block non-specialists after generation is complete. A low-cost Hostinger plan at roughly 7 euros per month was presented as enough for personal apps and lightweight SaaS projects, with Ubuntu as the only notable manual server choice. The broader significance is that AI coding tools are now expanding beyond code generation into provisioning, deployment and maintenance.

GrokBot pitches model-routing coworkers

GrokBot is being positioned as an always-on AI workspace built around persistent specialist agents rather than one-off chats. Its key differentiator is model routing: tasks can be sent automatically to the model deemed best suited, including external systems such as Claude Opus 5.5, while the user stays inside one thread. Connected integrations with Gmail, Google Docs, Slack, ClickUp, Firecrawl and Neon PostgreSQL are intended to let bots act on business tools instead of merely summarizing them. Voice interaction and a MacBook utility called Grok Bot Notch further push the product toward an ambient desktop-assistant model.

Microsoft 365 outage hits Copilot

A major Microsoft 365 outage disrupted Outlook, Exchange Online, SharePoint and Copilot across 74 countries for about 36 hours. Microsoft said the incident stemmed from an earlier network capacity upgrade, with maintenance later exposing a latent bug in the system that establishes secure connections. Automated health checks then amplified the failure by removing struggling servers from rotation, overloading those that remained and creating a cascading collapse. The episode underscores how deeply AI services such as Copilot now depend on the same cloud reliability chain as email and collaboration software.

Cheap open models raise cyber risk

New testing of AI-driven hacking found that closed frontier systems from OpenAI and Anthropic still perform best on complex attack chains, particularly when recovery and strategic adaptation matter. But the sharper warning came from cheaper open-weight Chinese models, including DeepSeek, which reportedly achieved most objectives in one network for under $3 per attempt. By contrast, a leading OpenAI model completed the same objectives at roughly $20 per run in the cited benchmark. That growing asymmetry means the most capable models may remain concentrated, while 'good enough' autonomous offense becomes broadly accessible at very low cost.

Videos covered

Previous briefings · AI