Anthropic pushed a cheaper agent model into market, while capital and policy both got more explicit numbers. MGX closed a $49 billion AI fund, and OpenAI floated a 5% US stake to ease political pressure.
Model releases & updates
- Anthropic launched Claude Sonnet 5, a cheaper option for running agents with safety improvements. That gives teams another model to compare on cost per task, not just raw capability.
- Anthropic’s Fable and Mythos models got global release after new safety testing changes. Broader availability usually means more enterprise rollout, especially where export rules had slowed adoption.
Money & moves
- MGX closed a $49 billion AI fund with OpenAI and Anthropic backer ties. That is one of the largest AI pools on record, and it keeps capital concentrated around frontier infrastructure and model bets.
- Etched hit a $5 billion valuation and $1 billion in sales for its AI chip. That is a real commercial wedge against Nvidia, not just a pitch deck story.
- AWS launched a new $1 billion AI unit built around embedded engineers. It signals that cloud vendors now sell implementation help, not just APIs and instances.
- Meta is exploring ways to sell excess AI compute. If it works, spare GPU capacity becomes a revenue line instead of dead weight.
Policy & governance
- OpenAI floated giving the Trump administration a 5% cut of the AI boom. That is a blunt attempt to lower political resistance to AI expansion, and it shows how financial terms are entering governance talks.
- Trump dropped restrictions on Anthropic’s Mythos and Fable models. The practical effect is fewer barriers on where those models can be deployed and sold.
- Midjourney is pushing Hollywood studios to disclose their AI usage in an ongoing legal fight. That could surface who is using generative tools behind the scenes, and on what terms.
- Anthropic said export controls on Claude Fable 5 and Mythos 5 were lifted. Less friction on exports usually means faster international distribution for frontier models.
Research
- Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification proposes a scalable safety-testing framework for agents. That matters because agent deployments need repeatable checks, not one-off red-team demos.
- Expert Evaluation of Clinical AI Tools on Real Point-of-Care Clinical Queries benchmarks tools on 620 real clinical questions. Real queries are harder than synthetic tests, so this is closer to what clinicians actually face.
- OSWorld2.0 expands computer-use benchmarking for long-horizon tasks. That raises the bar for agents that have to work across many steps, windows, and app states.
- HARC studies harmfulness and refusal directions for safety alignment. The practical angle is better control over when models refuse versus when they comply.
New marketing skills & frameworks
Fresh AI skills, agent skills, and frameworks a marketer can pick up this week — each linked to its source.
- Paid Media Audit With Live Data (ds-paid-audit): Connects directly to Google Ads, Meta, and LinkedIn without exporting data to run structured audits that detect budget waste, audience overlap, and creative fatigue. — DataSlayer
- AI-Optimized Answer Engine Strategy (AEO): A new competency focusing on optimizing content, images, and video for AI-generated answers in Google AI Overviews, ChatGPT, and Perplexity rather than traditional search links. — Clayton Johnson
- Vibe Coding for Marketing Agents: Enables marketers to build custom ROI calculators, scrapers, and autonomous marketing agents using AI code tools like Claude Code without writing traditional code. — YouTube
How this digest is made
This roundup is generated by Letaido, an AI agent that runs the whole pipeline automatically. Each week it pulls the latest posts from a curated set of AI-industry sources — major tech-press outlets, the AI labs’ own blogs, arXiv, and live web-event streams from Firehose taps — then uses AI to score every item for relevance and importance, drops near-duplicates, and ranks what made the cut. This issue was drawn from 701 stories across 10 sources; 18 made the final digest. A human editor reviews each issue after it publishes and their feedback tunes future editions. Sources are linked inline so you can read the primary reporting yourself.
