Good morning. There’s a fine line between prepping kids for the future and using them as the future’s focus group. One San Diego charter school seems determined to find it with two humanoid robots and a $500,000 price tag. The robots are Ameca, advanced humanoids that look like they wandered off the set of a very expensive sci-fi therapy session. The school calls it research into physical AI as a teaching partner, which sounds better than beta testing the future on children.
OpenAI cut inference costs by more than half. The AI industry has been chasing more Nvidia chips, but OpenAI just found extra capacity inside the servers it already has. Just to be safe, they tested the new optimizations on free ChatGPT accounts, and that slice of traffic only needed a couple hundred Nvidia GPUs. That’s almost nothing for a company trying to serve ChatGPT to everyone with a pulse. OpenAI has not said exactly what changed, but the savings seem to come from making models waste less work by reusing old calculations, batching questions, shrinking the math, and routing easier prompts to cheaper systems. Now OpenAI has options. Give users more GPT, cut API prices for developers, or keep the savings to make investors happy. Maybe now they’ll be able to compete against China’s open models. (The Information 🔑)
A fusion reactor just powered a lightbulb. It might not sound like a big deal, but it’s the first time electricity has been extracted directly from charged particles inside plasma. This is the point where fusion stops sounding like a science fair project and starts becoming an energy story. Most reactors lose a pile of power moving energy through heat, steam, and turbines. Direct conversion grabs electricity from the reaction before that detour. Realta Fusion, the startup behind the breakthrough, says it could capture about 90% of the energy, compared with 33% for steam turbines. In a commercial reactor, that extra power could feed back into the machine and help keep the plasma hot. Every energy revolution needs a lightbulb moment, and fusion finally found one. (TechCrunch)
Strands Agents is the open source agent harness SK from AWS, powering everything from production backends to physical robots to code that writes itself. Build production-ready agents in a few lines of code, with any model, anywhere.
- Context management, execution limits, and observability before you write a line of config
- Hooks to monitor, modify, and debug
- 100% agent accuracy with steering
👀 closer look
Anthropic unleashed an agent swarm for scientists. Claude Science is a new AI workbench built to run scientific workflows and speed up research. A lead agent splits work among specialist agents while a reviewer agent checks facts before anything reaches a paper. Now Claude sits inside the lab workflow, with access to more than 60 scientific databases, research tools, code, compute, figures, and audit trails. One researcher used it to map his field for $26. He fed it a library of 6,576 papers and a question he had been chewing on for years. Claude Science helped him explore the data and quickly uncover missing relationships. Anthropic’s aimed the first version at life sciences because pharma has the money, but the bigger play is every field buried under papers no one has time to read. (Anthropic, Forbes)
AWS wants to install your AI agents for you. It’s launching a new unit of Forward Deployed Engineers who go directly into businesses to build and roll out AI systems. AWS says the unit will start with thousands of engineers, with small pods embedded in customer companies. That is an easy upsell for anyone already running on AWS. Let Amazon build the agents, connect the workflows, and the next wave of spending flows through AWS cloud services and its Agent Marketplace. OpenAI and Anthropic launched similar deployment teams earlier this year, but AWS already has the customers. AI agents are becoming the new enterprise software, and Amazon brought the Geek Squad for Fortune 500s. (Reuters, AWS)
Did you know? Cloud services can cost 18-20x what self hosting does. It’s no wonder open source is so hot.
Claude Sonnet 5 makes AI agents cheap enough to matter. Anthropic is making Sonnet 5 the default model for free and Pro users, putting agentic AI in the cheap seats. The pitch is that it can plan work, use tools, write code, and run longer tasks that needed bigger models a few months ago. At launch, it costs $2 per million input tokens and $10 per million output tokens, below Opus 4.8, GPT 5.5, and Gemini 3.1 Pro. Sonnet 5 also scored close to Opus on agentic coding, giving developers a cheaper model they can run all day. The agent race is turning into a price war with tools. You can stop hiring the genius for every spreadsheet errand. (Anthropic)
fun stats
🛰️ $430,000. What Nvidia is paying engineers to help build AI data centers in space.
🪴 10%. Hiring growth at AI pilled companies over 2 years, according to Ramp and Revelop’s analysis of 21,559 firms. So much for the AI jobs apocalypse.
🕶️ $20. What Meta is going to charge smart glasses owners to bypass new monthly AI rate limits.
🛍️ 40%. Better conversion rates from Amazon’s ChatGPT ads compared to search, email, or social media during Prime Day.