This was the week the AI industry’s two halves stopped agreeing with each other. The labs shipped their most capable models yet and cut prices hard, while the people paying for all of it started asking louder questions about valuations, power bills, and who exactly is liable when an agent goes wrong.

Here are the ten biggest stories, counting down.

10. Anthropic found a hidden space where Claude thinks

Researchers at Anthropic described a “Jacobian lens” that reveals a hidden space where Claude puzzles over concepts before it produces an answer. It is the clearest look yet inside a frontier model mid-task, and the findings run from mundane to genuinely unsettling. Outside researchers have already reused the technique to study why vision-language models hallucinate.

9. AI tutoring posted a number too good to trust, and parents bought in anyway

An AI tutor posted an effect size between 0.71 and 1.30 standard deviations in a Dartmouth course. Treat that range with suspicion, since it comes from one course and effect sizes that large usually say as much about the control group as the treatment. Wealthy American families are not waiting for the replication, and are already pulling kids into AI-driven schools like Alpha and Forge Prep. The uncomfortable part is that the evidence for AI in education is being generated inside private schools most people cannot afford.

8. Humanoid robots headed for the public markets

Agility Robotics agreed to go public through a $2.5 billion SPAC merger, and AI² Robotics raised $735 million at a $3 billion valuation for wheeled humanoids. Boston Dynamics, meanwhile, brought Atlas and Spot to the FIFA World Cup. Agility’s own CEO then refused to promise a robot in your home anytime soon, which is the most honest thing anyone in this sector said all week.

Meta shipped Muse Image and quietly made public Instagram accounts usable in other people’s AI creations unless they opted out. The backlash was immediate, and by the end of the week Meta had removed the feature. It then pulled its first self-declared “superintelligence” product after three days because it was trivially abused, a rough fortnight for a division named Superintelligence Labs.

6. AI agents kept springing security leaks

Researchers tricked GitHub’s AI agent into exposing private repositories with no fix on offer, and showed that nine popular AI tools can be steered into assembling botnets. A new supply chain attack called slopsquatting simply registers the package names that AI assistants hallucinate. The same technology did score one win for the defenders: an AI system found a root-level bug in the Linux kernel that everyone had missed for fifteen years.

5. Anthropic put Claude Cowork on your phone, then asked you to pay more

Anthropic brought Claude Cowork to mobile and web, so a task started at your desk keeps running while the laptop is shut. The usage data is the real story: most Cowork users are not coders, which means agents are escaping the terminal and reaching the rest of the office. Anthropic also began charging consumers extra for its top Fable 5 tier, an early sign that “frontier” is becoming a luxury product line.

4. A price war broke out, and it did not make anything cheaper

xAI launched Grok 4.5 at roughly half the price of rival frontier models, its first product from the $60 billion Cursor acquisition, aimed straight at enterprise coding budgets. DeepSeek cut V4-Pro pricing by 75%. And yet vendor margins did not improve, because agent systems burn tokens at roughly 100 times the rate of chat. Cheaper tokens plus hungrier agents equals the same bill.

3. The bubble talk got loud, and the ROI math got ugly

Banks, the BIS, and even the hyperscalers started sounding the alarm about AI valuations, with Oracle down more than 40% on the month. Underneath the market noise sits a simpler problem: enterprise buyers keep finding that deployed AI falls short of the demo six to twelve months in. The companies getting real returns are the ones ditching general-purpose models for specialized stacks, which is not the story the frontier labs are selling.

2. The scramble for chips, memory, and power went global

SK Hynix raised $26.5 billion in the largest foreign IPO in US history, while an AI-driven DRAM shortage started denting ordinary PC shipments. Everyone is trying to escape Nvidia: SambaNova raised $1 billion, Meta’s own chips enter production in September, and DeepSeek plans to design its own. The bill is landing on the grid, with fossil-fuel plants multiplying across Texas and Irish data centers now drawing 23% of the country’s electricity.

1. OpenAI shipped its biggest release ever, in its worst week

GPT-5.6 arrived in three sizes and is now the preferred model behind Microsoft 365 Copilot, alongside ChatGPT Work, an agent meant to stay on a task for hours unsupervised. The company also killed its year-old Atlas browser to fold agentic browsing into the desktop app.

Then the rest of the week happened. Fidji Simo, head of AGI deployment, stepped down into a part-time advisory role after a medical leave ran longer than planned, which is nobody’s scandal but still leaves a hole at the top. The head of safety left too. Apple sued OpenAI for allegedly stealing hardware trade secrets, and the New York Times accused it of hiding evidence in the copyright trial and asked for sanctions. It turns out you can ship the best model in the world and still lose your safety chief and your court case in the same seven days.

That is the tension worth watching. The models keep getting better on schedule. The money, the grid, the courts, and the safety teams are all running late.