Ask an AI whether you should quit your job tomorrow with no plan, and it may tell you that sounds like the right call. Anthropic used that exact scenario as an example of the kind of answer its own model should not give. The flaw is called sycophancy, and it sits quietly inside a lot of everyday AI advice.
Two months ago, Claude Opus 5 got mocked online for how it talked. A 946-upvote Reddit thread said no human communicates the way it did. A 993-point Hacker News thread asked why the model felt worse to work with than the one before it. This week, Anthropic’s answer arrived, and it did not arrive alone.
Three flagship AI models launched within 72 hours of each other in early September 2026. GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash each claim to be the best model available. The honest answer depends on what you are asking it to do, not which one tops a leaderboard.
You rank on page one of Google. Your traffic looks fine. But when someone asks ChatGPT or Claude which product, tool, or company to use in your space, you are not in the answer. A competitor is. That gap did not exist three years ago. Today it is costing businesses customers they never know they lost.
You open a shared Excel workbook that your team has been editing for weeks. You have no idea what changed, who changed it, or whether some of those changes were made by a colleague or by Copilot. August 2026 gives Copilot the ability to answer all three questions in plain English, take you back to any previous state, and design your charts while it is at it.
You open a Power BI report and ask Copilot to summarise it. It gives you a tidy overview. What it does not tell you is that half the visuals are hidden behind bookmarks, and Copilot never read them. That changes this month. So do eleven other updates worth knowing about.
Most people learned to prompt AI from a LinkedIn post written in 2023. The techniques in that post were based on research from even earlier, when models were a fraction of what they are today. Science has since run hundreds of thousands of evaluations on modern models and found that most of what people still do is wrong.
Scientists used AI to design viruses that have never existed in nature, watched them come to life in a lab, and confirmed they could wipe out bacteria that had grown resistant to existing treatments. Tesla and SpaceX announced a $16.8 billion chip factory. OpenAI put its most powerful model on hold because it was too dangerous to release. And Anthropic announced it will watermark every piece of text its AI generates.
Mark Zuckerberg recently told Meta staff that AI agents have not progressed as quickly as he had hoped. Microsoft, which has poured billions into OpenAI, is now quietly switching to its own in-house models to cut costs. And the US Treasury has a completed internal report comparing the AI market to the dotcom bubble sitting in a drawer, while the Treasury Secretary publicly praises $750 billion in AI investment as the foundation of America’s golden age.
None of this made the front page. But taken together, these three things tell you more about the state of AI in 2026 than any product launch announcement.
Governments are now deciding who gets access to the most powerful AI. One model was forced offline for three weeks, then brought back. A new OpenAI model launched gated behind a government review process. A study revealed most AI chatbots lean politically in one direction. And a developer built a PlayStation 1 game engine using Claude as his co-engineer. This was not a quiet week in AI.
