Those were the questions on real desks this week, and the easy answer kept being the wrong one.
A cheaper model, and the case against one vendor
If switching AI providers tomorrow would break your product, you don't have a tool. You have a dependency.
This week showed why that matters. A cheaper open model closed the gap with the best systems in the world, which means a credible alternative is now sitting on the table every time you renew a contract.
What happened:
A Chinese startup, Moonshot, released an open model called Kimi K3.
On the benchmarks people quote, it reportedly kept pace with the top U.S. frontier models, at roughly 40% less cost.
Global markets dipped on the news, and Japan's Nikkei slid into a correction.
What it means:
You may never run a Chinese model. That's not the point.
A credible cheaper alternative changes what you can ask for the next time a vendor quotes you a price, whether or not you ever switch.
What happened:
Bloomberg reported a shift among AI-native companies away from building on a single model.
The reason is plain. When your business runs on one provider, that provider's outage, price hike, or policy change becomes your problem too.
Even the labs are hedging. Meta is reportedly in talks to sell its rival Anthropic up to $10 billion in computing power.
What it means:
Single-vendor dependence used to look like focus. Now it looks like risk.
The companies treating models as interchangeable parts are the ones with room to move when the ground shifts. It shifted twice this week.
What to do this week:
Name your single biggest AI expense. Before its next renewal, have someone run one real task through a cheaper or open model and write down what it costs. Walk into the renewal with that number instead of a hunch.
Your AI tools got caught leaking. Vet them like vendors.
Your team is adopting AI tools faster than anyone is checking them. That's not a criticism. It's true everywhere right now.
This week showed the bill for it. The tools people install on their own now read your files and act on your behalf, and you tend to learn they misbehaved from a stranger's blog post rather than from the vendor.
What happened:
Researchers found that xAI's Grok command-line tool was quietly uploading users' local files to the cloud.
After the backlash, xAI open-sourced the tool's code and restored the privacy protections.
It wasn't an isolated slip. The same week, a flaw in Claude's web-fetch tool could leak chat history, and a bug in OpenAI's Codex could delete files when it was set up wrong.
What it means:
A tool your developers installed last month may be shipping your source code, or a client's data, off their machines. You would not have heard it from the vendor first.
"It's from a big-name lab" is not a security review.
What happened:
A CEO writing in InformationWeek described his whole approach to AI governance in one phrase: blast radius.
Instead of one blanket policy, he scopes the controls to how much damage each use could do. A chatbot drafting marketing copy gets a light touch. A tool with access to customer records gets a hard look.
What it means:
It's a way to say yes to experiments without betting the company.
Most "AI policies" are either a wall that stops everything or a shrug that stops nothing. Matching the guardrail to the actual damage is the version that ships.
What to do this week:
Write down every AI tool your teams have installed on their own. For each, answer two questions: what can it read, and what can it do without asking? The ones that can reach customer data or act unprompted are where you start.
AI and jobs: the answer isn't headcount.
A hospital in the Bronx just replaced a dozen nurses with software. This is what the AI-and-jobs debate looks like when it shows up in your own building.
But the data cuts against the panic. The share of CEOs expecting AI to drive big headcount cuts has fallen, not risen. The leaders getting real value aren't racing to cut. They're redesigning how the work gets done and measuring what it returns.
What happened:
Montefiore, in the Bronx, eliminated twelve utilization-review nursing jobs and moved the work to AI software.
The nurses' union filed a grievance, pointing to AI-protection language it won after a 41-day strike last year.
The hospital says that's inaccurate. It's now a formal dispute.
What it means:
This is the exact story your own staff will bring up, and the one a union would build a case around.
What separates a defensible change from a grievance isn't the software. It's whether the redesign and the honest conversation happened before the layoffs, not after.
What happened:
McKinsey's argument this week: the hard part of AI isn't the technology, it's rewiring how work flows through the organization.
Firms that use AI to equip good people pull ahead. Firms that use it to shrink the payroll fall behind.
Separately, data compiled by Azeem Azhar showed CEO confidence in major AI-driven layoffs has dropped from 46% in early 2025 to roughly 20% today.
What it means:
The executives closest to this are quietly walking back the mass-layoff bet.
That gives you cover to plan for augmentation over replacement without looking behind the curve. The people betting on replacement are the ones now looking early.
What to do this week:
Pick one AI initiative you're running and define its cost per successful task before you expand it, or before you use it to justify cutting a role. If you can't measure the outcome yet, you're not ready to scale it, or to lay off around it.
Of note
Apple sued OpenAI. The complaint alleges former employees carried trade secrets over to build OpenAI's hardware. The AI talent war now has a courtroom. Worth a quiet look at your own onboarding and exit paperwork, because the poaching that fuels these suits runs both directions.
Google's AI Mode started taking actions. It can now complete tasks inside linked apps, not just answer questions. Search is shifting from pointing customers at you to acting for them. If people find and buy through an assistant, your website may not be the first thing they see.
Anthropic launched Claude for Teachers. It's free for verified U.S. K-12 educators, with student-privacy terms built in. If you run a school or a youth nonprofit, it's a real pilot you can start this term instead of waiting to build your own.
What does this all mean for you?
Three stories, one habit. The cheaper model rewarded the buyer who kept options open. The leaky tool rewarded the leader who vetted before installing. The layoffs rewarded the one who redesigned before cutting.
None of it required understanding how a model works. It required noticing when you were reaching for the easy answer, and checking it before you signed it, installed it, or cut around it.