The most useful development this week is not a flashy launch. It is the steady march of open-weight models you can now run yourself, which matters far more to a business owner worried about where their data ends up than the next frontier benchmark.
📌 A wave of open-weight releases over the past six weeks now rivals the big commercial models, with Z.ai's GLM-5.1, Moonshot's Kimi K2.6, DeepSeek V4 and Alibaba's Qwen 3.6 mostly shipping under permissive licences at a fraction of frontier pricing. The real shift is that these are now a genuine option for running in-house on relatively inexpensive hardware, not just in a data centre. That means a business can keep its data entirely on its own machines, and it is worth asking any vendor selling you an AI feature what model sits underneath it and where your data goes.
📌 Anthropic released Opus 4.8 on Wednesday, its most advanced public model, at the same price as the version it replaces. I spent more time with it this week and it is noticeably better at flagging its own hallucinations, which is the upgrade that actually counts. For business use a confident wrong answer is the expensive failure mode, so a model that tells you when it is unsure is worth more than another point on a benchmark.
📌 Google flipped its core search engine over to Gemini 3.5 Flash, with the traditional search box essentially retired in favour of a conversational interface by default. This is an enormous change for search, and it pushes organic results and paid ads even further down the page. The practical upshot is that the value of Google for a small business keeps shrinking, and if you have been relying on showing up in search results to bring in leads, that channel is quietly being closed off.
📌 Microsoft pushed a redesign of its Office Copilot tools and moved its computer-using agents to general availability in Copilot Studio. This is a good step forward for Microsoft, but they are still a long way behind the more serious agents like Claude Cowork, Perplexity Computer and Manus. If you want an agent that can actually work across your systems today, Microsoft is not yet where I would look first.
📌 The open-source agent framework OpenClaw has logged at least 454 vulnerabilities, with Gartner advising enterprises to block downloads. The good news is there are safer options. Hermes Agent from Nous Research is a more secure, MIT-licensed alternative, and it is honestly pretty easy to stand up your own agent framework in Python these days, with lightweight tools like Nanobot running in only a few thousand lines of code. The point is that you do not have to accept a framework with a security record like that one.
📌 Sam Altman told Commonwealth Bank CEO Matt Comyn he was "pretty wrong" about AI quickly wiping out white collar jobs, walking back his earlier predictions. I share his view. AI is a useful tool that in certain cases can genuinely lift productivity, but it is a long, long way from replacing workers, and it needs constant close supervision to be trusted with anything that matters.