This Week’s News

1. OpenAI updates ChatGPT with a more factual and focused GPT-5.6 Sol for Plus and Pro users, including a slider for choosing how much reasoning to apply, while making GPT-5.6 Luna the default for Free and Go users. [Link]

2. Meta releases Muse Glimmer, an open-weight model for running multimodal, multi-step agents on a single consumer GPU. The 30-billion-parameter model supports more than 100 languages and can use tools, work with screenshots, and recover from failed actions locally. [Link]

3. Nvidia releases Nemotron 3.5 Lightning for high-volume agent tasks and NeMo Switchyard for routing workflows between models. Nvidia says the open tools can complete tasks faster and reserve expensive frontier models for requests that need them. [Link]

4. Google says the Gemini app has reached one billion monthly users, while OpenAI says ChatGPT passed one billion weekly users in July. Although the figures are not directly comparable, Gemini’s rapid growth suggests the gap in consumer AI is narrowing. [Link]

5. OpenAI slows the development and testing of Astra after the model reaches a critical cybersecurity threshold. Preliminary evaluations suggested it might independently identify and carry out attacks against well-protected real-world systems. [Link]

6. Anthropic and Andon Labs find that Claude Fable 5 can control a drone to locate and follow a person, but faulty scene reconstruction prevents reliable end-to-end navigation. [Link]

7. Cloudflare launches Kitesurf, a cloud-hosted browser built specifically for AI agents. The company says it uses three to seven times less CPU or memory than Chromium for common tasks, although it is slower and does not yet support video or WebGL. [Link]

8. Meta launches Muse Code, its first coding agent, alongside the Muse Spark 1.2 model. The preview can plan changes, write code, validate results, and coordinate fleets of agents. [Link]

9. Anthropic confirms it is building a custom-silicon team to co-design chips and AI models so Claude can run faster and more efficiently. [Link]

10. Stanford researchers use generative AI to design complete viral genomes, producing 16 functional bacteriophages capable of killing E. coli. The work could contribute to treatments for antibiotic-resistant infections but raises urgent biosafety concerns. [Link]

11. Google DeepMind’s WeatherNext AI predicts tropical cyclones with roughly one extra day of useful lead time compared with existing models. DeepMind is open-sourcing its hurricane-season models, while forecasters caution that human expertise remains essential. [Link]

12. Microsoft Research introduces CARE-X, a vision-language model combining chest X-ray interpretation with diagnostic scores and spatial grounding. The retrospective research system reached 94% accuracy on the ReXVQA benchmark but is not approved for patient care. [Link]

13. Researchers demonstrate an AI scientist that can set up and run a simulated X-ray-scattering experiment using equipment at the SLAC National Accelerator Laboratory, adapting its plan after detecting a motor fault. [Link]

14. A study involving 1,682 adults finds that ChatGPT-generated short stories are rated more absorbing and higher quality than human-written stories, despite readers preferring work they were told had human authors. [Link]

15. Spotify will label fictional AI-generated artists and exclude their music from recommendations unless a user explicitly follows them. The “AI Persona” badges begin appearing in September. [Link]

16. Anthropic says it will add machine-readable watermarks to Claude-generated text and C2PA credentials to generated files to comply with the EU AI Act’s transparency rules. [Link]

17. The White House finalises a voluntary framework for testing advanced AI models but reportedly plans to keep its criteria private and share them only with selected technology companies. [Link]

18. Optical-networking startup Lumilens emerges from stealth with 900 million USD in funding and a 5.51 billion USD valuation. The company says it has billions of dollars in orders for optical chips connecting large AI computing clusters. [Link]

19. River AI raises 1.1 billion USD to build personally trainable assistants, beginning with an API for fine-tuning open models using reinforcement learning and LoRA without managing infrastructure. [Link]

20. Amazon-owned Zoox begins charging for rides in its purpose-built, steering-wheel-free robotaxis in Las Vegas, moving from public testing into commercial service. [Link]