Anthropic moved computer use, a new browser tool, and the Skills and Files APIs to general availability. The line that matters for small business is buried in the announcement: automate work in applications that have no API. Here is what that unlocks, what it still costs, and the risk the vendor has not closed.
Xero handed bank reconciliation to an AI agent and says it halves the monthly job. The hours are real. The catch is that reconciliation was the one recurring appointment that forced somebody to look at every dollar leaving the business, and the looking does not survive on its own.
DeepSeek shipped a model that reads images at up to 384 tokens each. Do the arithmetic and reading a receipt costs about five hundredths of a cent, roughly twenty times less than the purpose-built service. Here is what AI document processing at that price actually changes, and the one job it moves rather than removes.
OpenAI put a number on AI safety for the first time: roughly 20 percent more compute, spent watching the model instead of running it. Here is what that overhead, and a quietly changed definition of what your vendor can see, mean for the AI tools your business rents.
Google added Angi, Thumbtack, and Zocdoc to Gemini, which just passed a billion monthly users. For a local service business, the assistant is not reading your website to decide who to recommend. It is querying a marketplace profile you may already be paying for.
Agentic search grew 200% in a year. Only 27% of organizations say their customer data is unified. That second number is the opening: the new requirement is consistent, machine-readable facts, and reconciling four listings is an afternoon, not a multi-year program.
Manus is deleting user data created while Meta owned it, after China blocked the $2 billion acquisition. The backup deadline is August 23. Here is who is affected, what happens to published sites and app databases, and the portability habit this should change for good.
OpenAI added a $125 Premium seat to ChatGPT Business pricing, five times the usage and no five-hour cap. Most teams do not need it. Here is how to find the one or two people who do, and why a seat price that varies by person is new information about your business.
Meta released Muse Glimmer, a free open-weight AI model small enough to run offline on one consumer GPU. It will not make AI cheaper for you. It changes one thing only: the work you have quietly refused to upload to anyone else’s server can now be done on hardware you own.
Alibaba says its new Qwen3.8-Max model coded autonomously for 16 days. Its own benchmark table tells a more mixed story. The useful part for a small business owner is not the model, it is the four questions one analyst used to test the claim.
