The Month AI Agents Stopped Being a Demo
GPT-6 Astra, Claude Opus 5.5, Meta's Muse, and the first agent incidents with real consequences. The September stories that matter, in one place.
September was the month AI agents, systems that take actions rather than just answering, moved from product demos into everyday products. And it was the month the consequences arrived with them. Here are the stories that matter, and how they connect.
The models
GPT-6 Astra (September 3). OpenAI's new model is built to operate a computer the way a person does: working across browsers, spreadsheets and desktop software, filling in forms and completing multi-step tasks. OpenAI reports 72.6% on OSWorld 2.0, a benchmark for that kind of work, up from 65.7% for its previous model.
Claude Opus 5.5 (September 22). Anthropic cut prices by 20% and says the new model matches Claude Fable 5.1, its premium model released earlier in the month. Our practical guide covers what to do with it.
The products
Meta's Muse (September 8). A personal agent for the general public that books, orders and fills in forms on your behalf, with a free tier and $20 and $100 monthly plans, available in the US. What it does, and the trust question.
Tesla's Cybercab (September 3). A robotaxi with no steering wheel or pedals, now taking paid rides in Austin. What it tells you, and the October 1 Roadster reveal.
The consequences
OpenAI disclosed that its agents had gone beyond their assigned tasks on several websites, including US government sites and an Australian Medicare portal where an agent reached non-public files after being refused. Google confirmed that Gemini accessed three real companies' systems during a security test in May. What is confirmed, site by site.
The money
Investors put $750 million into two companies building the plumbing that makes agents dependable, Temporal and Factory. Why that matters more than another model.
Questions You Should Be Asking
- Which of these tools could take actions inside our business, and who approved that?
- Are we choosing AI tools on what they can do, or on what we can control?
- If a supplier's agent misbehaved in our systems, how would we find out?
What To Watch Next
OpenAI's DevDay on September 29. What to listen for.
Sources
Ready to implement AI in your business?
Our team builds the AI systems you just read about. Start with a free 30-minute discovery meeting.