The week made one thing harder to ignore: model capability is getting cheaper faster than companies, regulators, grids, and managers can absorb it.
The headline version of the week was easy to write.
Anthropic shipped a faster Claude and kept a scarier one locked down. OpenAI pushed deeper into chips, drug discovery, office agents, and the next flagship model. DeepSeek V4 and other Chinese open-weight releases pulled the price floor down again. Google gave developers more credible local models. Mistral put serious open-weight capability within reach of teams that want to self-host.
That is the pleasant version: more capability, more choice, lower cost.
The useful version is less tidy. This was the week the cheap-intelligence story ran into expensive reality.
Start with the buyers. If you run software, legal, finance, HR, healthcare, or education, the model menu got better. But the number of decisions you have to own grew faster. You need to know whether an AI hiring screen has EU paperwork attached. You need to know which OAuth grants employees handed to AI plugins. You need to know whether the agent touching a credit card can place an order, dispute a charge, or leak a customer's data on the way there.
The agent is no longer a side tool. It is becoming an actor in the workflow.
That explains why the security stories mattered. Singapore telling banks to patch like a model is hunting them, the AI plugin that leaked production, Anthropic's contractor exposure, and Nebraska's line on fabricated citations all point to one buyer problem: AI mistakes are moving from private embarrassment to operational evidence.
If a model invents a citation in court, it is not a funny screenshot. If an agent leaks production access, it is an incident. If a hiring algorithm lacks documentation after the EU deadline, it is a compliance problem. The organization cannot hide behind "the tool did it" for much longer.
Then there is the grid.
Oracle's power posture, Anthropic's chip dependence, Google splitting its AI chip strategy, Marvell and Broadcom moving around TPU economics, data centers in orbit, and repeated stories about cables, storage, cooling, and transmission all said the same thing in different accents. Inference prices can fall. Infrastructure timelines do not obey the same curve.
That gap matters for planning. A product team can swap models in a sprint. A utility cannot build a substation that way. The companies with locked-in power, custom chips, and data-center sites get a different cost curve from everyone buying capacity after the rush starts.
Labor had the same split.
Apple teaching engineers AI as a baseline skill, Snap saying AI writes 65% of its code, Microsoft offering a golden handshake, NEC putting 30,000 employees on Claude, and law schools making AI training explicit all point to the same workplace rule: AI fluency is moving from optional advantage to required literacy.
But fluency is not the same as safety. Half of companies still had no AI governance. Schools were behind students. Courts were still dealing with fake citations. The practical gap was not access to tools. It was the absence of adult supervision around them.
For buyers, the move is boring and urgent: write down where agents can read, write, spend, approve, and escalate. Audit OAuth grants. Run a real model comparison on your own tasks. Put human verification where money, legal exposure, hiring, clinical judgment, or customer trust changes hands.
For sellers, the empty pitch is "AI-native transformation." The useful pitch is a control surface. Show the audit log. Show the permission boundary. Show the fallback model. Show what happens when the agent is wrong, slow, blocked, or too expensive.
This is where the week landed. Intelligence got cheaper. Accountability did not.
The week in one line: the model is becoming the easy part, and the operating wrapper around it is where the real work starts.