1. Google Launches Nano Banana 2 Image Model
Google released Nano Banana 2 (Gemini 3.1 Flash Image), an upgraded image model that is significantly faster and cheaper than its predecessor. It supports 4K resolution, improved text rendering, and consistent character generation for up to five subjects, taking the top spot on major text-to-image leaderboards.
2. Anthropic Rejects Pentagon Demand for Unrestricted Model Access
Anthropic CEO Dario Amodei rejected the Pentagon's request for unrestricted access to Claude, citing concerns over mass domestic surveillance and lethal autonomous weapons. While rivals like OpenAI and xAI have reportedly accepted similar terms, Anthropic maintains that these use cases undermine democratic values.
3. Claude Code Updates: Auto-Memory and Planning Reviews
Anthropic introduced auto-memory for Claude Code, allowing the agent to persist project context and preferences across sessions via MEMORY.md and CLAUDE.md files. Additionally, new community-driven commands like /plan-exit-review enable developers to force the agent to critique its implementation plans before writing code.
4. OpenAI Realtime API Reaches General Availability
OpenAI announced the general availability of its Realtime API and the gpt-realtime model, which offers improved speech-to-speech capabilities. The update includes better instruction following, lower latency, and a new prompting guide specifically for voice-based systems.
5. Cursor Agents Gain Virtual Machine and Desktop Control
Cursor upgraded its cloud-based agents with dedicated virtual machines and desktop control, allowing them to build, test, and validate code autonomously. This capability enables agents to verify their own work within a sandbox before submitting pull requests.
6. OpenAI Hires Former Meta and Apple Infra Lead Ruoming Pang
OpenAI hired Ruoming Pang, a former AI infrastructure lead at Meta and Apple, to join its research team. Pang previously led Apple’s models group and was a key figure in the development of Apple Intelligence and Meta's Superintelligence Labs.
7. Nous Research Open-Sources Hermes Agent
Nous Research open-sourced Hermes Agent, a personal AI agent designed to live on messaging platforms like Telegram, Slack, and Discord. The agent features over 40 built-in tools and is designed to learn and build reusable skills over time.
8. Google AI Edge Gallery Expands to iOS with FunctionGemma
Google expanded its AI Edge Gallery to iOS, showcasing the FunctionGemma model which performs on-device function calling with only 270M parameters. The app includes benchmarking tools to measure LiteRT performance across mobile CPUs and GPUs.
9. xAI Co-Founder Departs as SpaceX Merger Proceeds
xAI co-founder Toby Pohlen has departed the company, marking the seventh co-founder to leave since its inception. The move follows xAI’s recent merger with SpaceX, which is reportedly preparing for a major initial public offering.
10. Perplexity Supplies APIs to Major Android OEM
Perplexity is now providing its research and retrieval APIs to a major Android device manufacturer to power on-device AI features. This partnership aims to integrate deep search and summarization capabilities directly into the mobile handset ecosystem.
11. QuiverAI Launches Arrow 1.0 SVG Generation Model
QuiverAI emerged from stealth to launch Arrow 1.0, a specialized model for SVG generation. The model currently holds the top spot on Design Arena's SVG leaderboard and is now available in public beta.
12. Claude Free Plan Gains 150+ Third-Party Connectors
Anthropic added over 150 third-party connectors to Claude’s free plan, including integrations for Slack, Figma, and Asana. This update allows users to bring external data directly into their chat workflows without a paid subscription.
13. Hyperscaler AI Capex Projected to Reach $770B in 2026
A report from Epoch AI indicates that hyperscaler capital expenditure has quadrupled since the release of GPT-4, with total AI investment nearing $500 billion in 2025. Alphabet, Amazon, Meta, Microsoft, and Oracle are projected to spend a combined $770 billion on infrastructure in 2026.
14. Scrapling Framework for Scalable Web Scraping
Scrapling is a new Python-based web scraping framework designed to handle large-scale crawls and automatically adapt to website changes. It includes features to bypass common anti-bot systems and manage complex request logic with minimal code.
15. New Frameworks for Agentic Engineering Patterns
Simon Willison and other industry leaders are advocating for new agentic engineering patterns, such as using 'deep modules' to hide complex logic behind simple interfaces. These patterns aim to make codebases more navigable for AI agents as the cost of writing code drops toward zero.