Open Source AI Is as Good as Claude and Free. So Why Can't Companies Switch?
Read time: 5 minutes
Welcome to AI-Empowered Leaders. In this weekly email, I share actionable advice on AI adoption, use cases & strategic thinking from my experience as AI Trainer, Leadership Coach, and Consultant.
From the Trenches
For the last 12 months I've watched the same thing happen inside companies.
And to individuals too, myself included.
We lean on Claude and ChatGPT a little more each week. One more task. One more workflow.
It feels productive. It is productive.
But it behaves like a dependency. Quietly, you stop being able to work without it. I have written about this extensively before.
If you're a leader, this part will make you uncomfortable.
The companies you depend on have hungry investors and real pricing power. What happens when Anthropic or OpenAI decides tokens cost more, simply because they can?
And there's the constraint that never goes away: Every prompt, every uploaded doc, every Slack thread you let the AI read is company data. Even if they never train on it, you are handing your context to someone else's server.
For years, the obvious answer to all of this ("just run AI locally") wasn't real. The open models weren't good enough to matter.
That just changed.
The Briefing
Two things happened this month. They look unrelated. They're the same story from opposite ends.
1) The best open-source model yet shipped.
Z.ai released GLM-5.2 on June 13. Open weights, MIT license, free to download and run on your own infrastructure.
The specs are serious: 744 billion parameters, a 1-million-token context window, and pricing around $1.40 per million input tokens versus the frontier's much higher rates. On a hard coding benchmark (Terminal-Bench 2.1) it scored 81 against Claude Opus 4.8's 85.
Not better than the frontier. But close. At a fraction of the cost.
2) Anthropic planted a flag in your Slack.
On June 23, Anthropic launched Claude Tag. One shared @Claude that lives inside a Slack channel, builds context over time, and works on its own.
The detail that matters: Anthropic says 65% of its own product team's code now flows through the internal version of this tool.
This isn't a chatbot in a tab anymore. It's a teammate sitting in the room where your decisions get made.
The Real Story
Most coverage frames these as two cool launches. Here's what's actually going on.
The engine got cheap. The car did not.
Think of the AI model as an engine.
GLM-5.2 proves the engine is becoming a commodity. A near-frontier engine, 98% cheaper, that you can drop on your own servers.
But nobody drives an engine. You drive a car.
The car is everything wrapped around the engine: the chassis, the transmission, the dashboard, the wiring that connects it to the road. In AI, that's the harness. Your context, your memory, your tool calls, your integrations, the routing logic that sends each task to the right model.
That's where all the work lives. And it's car-specific. You can't pull the engine out of one car and bolt it into another and expect it to just drive.
The Lindy team learned this the hard way. When they moved off Claude to a cheaper open architecture, they couldn't lift and shift their prompts, their memory, their tool calls. They had to rebuild the whole car around the new engine. Publicly. From scratch.
So when a vendor tells you "switch and save 98%," remember what you're actually switching.
"Just as good" depends entirely on your work.
Open models are genuinely excellent at so called center-of-distribution tasks. The fat middle of knowledge work: a standard deck, a brochure site, first-pass copy, routine synthesis, familiar coding problems. Tasks with a thousand prior examples and an output a human can check in thirty seconds.
On that work, GLM-5.2 is fast, cheap, and sometimes better than the frontier.
Edge-of-distribution work is different. Novel, ambiguous, high-stakes, the things with no template. That's where you still want a frontier model.
The real question isn't "which model is best" but what does the distribution of my company's tasks actually look like?
Nobody ever sits down and answers that. And until you do, you cannot price the switch.
The even bigger trap is: Who owns your context?
This is the part my clients worry most about.
Claude Tag is brilliant. That is precisely what makes it dangerous.
It sits in your Slack, absorbs the messy institutional knowledge nobody ever wrote down, and becomes the thing your team can't imagine working without. Within privacy limits, it learns your business better than any document ever captured it.
Now do the math. Even if the open engine is 98% cheaper and nearly as good, are you going to rip out the one model that already knows how your company actually works?
We've taught leaders for decades that data is your edge.
So why are we suddenly comfortable renting our own context back from a frontier vendor?
Even with a flawless privacy policy, you are building your company's brain inside someone else's product. And that brain does not come with you when you leave.
The Playbook
Here's what you can do.
- Map your task distribution. Take a real sample of how your team uses AI and sort it into "center" (routine, templated, easy to check) and "edge" (novel, high-stakes, ambiguous). You can't make a model decision without this.
- Put a number on two things. Your monthly token spend, and your switching cost. The first is what you'd save. The second is what it'd actually cost to rebuild the car. Most leaders only ever calculate the first.
- Separate the engine from the car in your own stack. Don't hardcode your workflows to one vendor's model. Build so you can change engines later without scrapping the car. That optionality is worth more than this month's discount.
- Decide what context you will never hand over. Before you connect the next tool or channel, name the data that stays in-house. Make it a policy, not an afterthought.
- Pilot one open model on center-of-distribution work. Pick a low-risk, high-volume task type. Run GLM-5.2 (or similar) against it. Measure quality and cost honestly. This is how you find out where cheap actually works for you, instead of guessing.
The Monday Test
This week, try this: pull your last 20 AI tasks and tag each one "center" or "edge."
If most are center, you're paying frontier prices for an engine you don't need.
If most are edge, stop obsessing over the token price and start protecting your context.
Either way, you'll know something about your business that almost none of your competitors do.
My Honest Take: Claude Tag
It's the most strategically clever product Anthropic has shipped. Not because of what it does, but because of what it quietly accumulates. The convenience is real. So is the lock-in. Adopt it with your eyes open, and decide on purpose which channels it gets to live in.
Whenever you’re ready, here’s how I can help you win with AI:
1) AI Business Advisory
Spot, plan & launch AI use cases that save hours and unlock new value.
2) AI Enablement & Coaching
Take your team on a journey from AI beginners to critical-thinking power-users: working securely across tools, saving costs, and driving results.
I’ve already trained and coached 3,000+ leaders who are saving hours and performing at a higher level. Your team could be next.
Have questions? Hit reply to this email and I'll help out!
Talk soon,
Alex