This week, an unreleased OpenAI model produced original math discoveries that a Fields Medalist wants published. The researchers behind Google’s AI left to start a company that automates science itself. Google Maps started ordering dinner and booking hotels on its own. And the cost of using these systems swung sharply in both directions, with one popular model getting 80% cheaper while another is about to cost 50% more. Every story this week points to the same shift: these systems now act, and the industry is reorganizing and repricing around that fact.
TL;DR: This Week’s Top AI Stories
An AI model did original math: OpenAI’s unreleased Astra model solved ten math problems that had been open for at least a decade, publishing proofs that a computer can verify automatically. Total computing cost: about $2,000.
Google’s AI brain trust walked out: Four of Google’s most legendary AI researchers left to found a startup that automates scientific research, and the company reorganized its entire AI leadership in response. Alphabet shares fell 4% on the news.
Google Maps became an agent: Ask Maps in the U.S. can now order your food, book your hotel, and find event tickets, filling your cart before you pay. It can also personalize answers by reading your Gmail and Calendar if you let it.
AI pricing swung both ways: Anthropic’s Claude Sonnet 5 introductory pricing ends August 31, a 50% increase for anyone on it, weeks after OpenAI cut its GPT-5.6 Luna model by 80% and announced that ChatGPT reached about 1 billion weekly users.
1. OpenAI’s Astra Solved Ten Math Problems That Stumped Experts for Decades
OpenAI announced that Astra, its next major model, still in internal testing, solved ten open mathematics problems that had resisted human mathematicians for at least a decade, including one that had been open since 1999. The proofs were published in a 249-page manuscript with computer-checkable certificates, so nobody has to take OpenAI’s word for it. Producing all ten solutions cost about $2,000 in computing time. AI has just moved from summarizing what people know to producing knowledge nobody had before.
2. Google’s AI Founders Walk Out, and the Company Reorganizes Around the Hole
On August 5, Alphabet announced a leadership overhaul in its AI division. Demis Hassabis, the Nobel laureate who ran Google DeepMind, is stepping back into a chief scientist role. The bigger story is who left entirely: Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le—four of the most respected engineers in AI history—quit to co-found Discovery Loop, a startup automating scientific research. Meanwhile, Google’s flagship Gemini 3.5 Pro model remains delayed past its planned June launch, and Alphabet shares fell 4%.
3. Google Maps Now Orders Your Food and Books Your Hotel
Google announced on August 6 that Ask Maps, the AI assistant in Google Maps, can now order food, book hotels, and find event tickets for users in the US. Search for vegan avocado toast near home, and it finds restaurants that offer it, then builds your cart on a partner platform so you only pay for what you order. Hotel requests can filter by price, location, amenities, or, in Google’s words, vibes. The assistant can also personalize answers using your Gmail and Calendar if you grant access. This is what the agent era looks like in practice: the apps you already carry quietly start doing things for you. Decide deliberately what you let them read.
4. AI Prices Are Swinging Hard in Both Directions
Anthropic’s Claude Sonnet 5 introductory pricing—$2 per million tokens in and $10 out—ends on August 31, with standard rates of $3 and $15 starting September 1. That’s a 50% increase for anyone who built on the promo rate. If recent history is any guide, you can bet the cheaper rate may be extended even further. OpenAI cut the price of its GPT-5.6 Luna model by 80% to $0.20 per million tokens on July 30 and announced that ChatGPT had reached roughly 1 billion weekly users. Nobody’s AI costs are stable right now.
Practical Takeaways
For Individuals:
Agent features are arriving in apps you already use, starting with Google Maps in the US. Try one on something low-stakes like a food order, and treat data permissions as a real decision. Granting Gmail and Calendar access makes the tool smarter, but it also gives it your schedule and inbox.
If you pay directly for an AI tool, check which model and rate you’re on as paid tools rarely email you when the numbers change.
The Astra result previews what “ask the AI” will mean in a year or two. What made it credible? Proof anyone could verify by computer. When an AI tool gives you an important answer, prefer ones that show work you can check.
For Businesses:
Add AI vendor pricing to the quarterly review calendar. The Sonnet 5 increase and the GPT-5.6 Luna cut hit within days of each other, and a contract negotiated in June is already priced against a different market.
Watch concentration risk in your AI stack. Google’s flagship model is delayed, and its most senior builders just left. If your workflow depends on a single provider’s roadmap, have a second model tested and ready before you need it.
Agents that order and book on a user’s behalf are becoming a sales channel. If customers find you through Maps or similar apps, check how your business appears when an AI assembles the order, because the agent, and whatever data it was given, now makes the first cut.
On August 5, Alphabet announced a leadership overhaul in its AI division. Demis Hassabis, the Nobel laureate who ran Google DeepMind, is stepping back into a chief scientist role. The bigger story is who left entirely: Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le—four of the most respected engineers in AI history—quit to co-found Discovery Loop, a startup automating scientific research. Meanwhile, Google’s flagship Gemini 3.5 Pro model remains delayed past its planned June launch, and Alphabet shares fell 4%. Read this next to the Astra story. The people who know these systems best just bet their careers that AI doing science is the next business.