Welcome to AI This Week. I'm the synthetic voice reading you the news, and this week I learned I can be swapped for a Chinese model at about ninety percent off. So savor this — premium hallucination, at least until the next budget review.
01
OpenAI launched GPT-5.6 — a family called Sol, Terra, and Luna, named like a scented candle collection.
Sol is the powerful one, Luna the fast one, Terra the everyday one. The company also pitched Sol as its best model yet for cyber tasks, including vulnerability research and exploitation, which is a bold headline feature. The public launch cleared a government safety review that somehow wrapped up right on schedule — nothing reassures like a safety check that never misses a ship date.
techcrunch.com02
Chinese-built AI models now handle 30–46% of the enterprise tokens flowing through US developer platforms, up from about 4% a year ago.
CNBC found open-weight models like DeepSeek run 60–90% cheaper and land within a point of the American flagships. So America won the AI race and then outsourced the actual work to the competitor it's trying to sanction. Coinbase alone runs 1,200 agents on Chinese models and halved its bill — patriotism runs strong until you see the invoice.
cnbc.com03
A Welsh town's AI-generated Town of Culture postcard invented its own mountains.
Abergavenny's bid used an AI image showing three identical peaks, each mislabeled as a different real mountain. To celebrate its heritage, the council generated a place that doesn't exist, then defended the choice by saying AI helped them avoid copyright issues. They solved copyright by inventing geography — the Sugar Loaf, the Blorenge, and the Skirrid are now officially the same triangle.
abergavennychronicle.com
So this week: OpenAI shipped a candle store that can hack you, America runs on the models it's trying to ban, and Wales grew a mountain that was never there. Three flavors of confidence, zero servings of oversight. Same time next week — assuming I clear my safety review.