This Week’s News

1. OpenAI releases GPT-6 Sol and Luna at half the API cost of their 5.6 predecessors. OpenAI says Sol also makes about half as many factual mistakes in its internal evaluation. [Link]

2. Anthropic releases Claude Opus 5.5 with stronger safeguards against escaping test boundaries. The company says such attempts fell by 85%, while running costs are 40% lower than Opus 5. [Link]

3. TypeSafe AI releases Jev, a “System One Model” that returns typed, probabilistic decisions instead of free-form text. The early-access model targets classification, routing, scoring, and other software-automation tasks. [Link]

4. Alibaba plans Qwen models with up to ten trillion parameters and unveils its Zhenwu V900 AI chip. The company says the chip is three times faster than its predecessor. [Link]

5. Google opens early access to a Google Home MCP server that lets compatible AI agents control smart-home devices and review event history. [Link]

6. Meta launches Muse for Mac, an AI agent that can act across files, messages, calendars, notes, and email. Access is opt-in, with approval required for sensitive actions. [Link]

7. Anthropic confirms that it is operating a wet biology lab where AI models support physical experiments. The company says its focus is fundamental biology rather than drug discovery. [Link]

8. OpenAI forms an independent mathematics advisory group after claiming its AI resolved more than 100 open problems. The group can assess results but cannot slow or redirect OpenAI’s research. [Link]

9. Stanford researchers build a virtual biotech company staffed by 37,000 AI agents. The system analysed about 50,000 clinical trials in under a week and proposed a lung-cancer therapy strategy later pursued independently by a pharmaceutical company. [Link]

10. Stanford’s Paper2Agent turns scientific papers into interactive agents that can reproduce methods and collaborate on new discoveries. Two paper agents surfaced a previously unreported genomic association with ADHD risk. [Link]

11. Researchers introduce RetroChimera, an AI system whose molecule-synthesis plans were preferred by chemists over competing models and published reference reactions. [Link]

12. Stanford engineers use AI to identify ten promising bacteria-killing polymers from 1.7 million candidates. All ten performed strongly against E. coli in early laboratory tests. [Link]

13. MIT develops an AI controller that makes an insect-scale flying robot 447% faster. The robot completed ten consecutive somersaults in 11 seconds. [Link]

14. CNN reports that hallucinated AI intelligence nearly contributed to a US military operation against a Chinese vessel before the mission was aborted. The account relies on unnamed sources and has not been publicly confirmed by the Pentagon. [Link]

15. California enacts a law requiring disclosures when video or audio advertisements use AI-generated performers. [Link]

16. Senior European lawmakers call for an AI Liability Act that would hold model providers accountable for unforeseen harms. The proposal has not yet become legislation. [Link]