Companies Are Firing Their AI Vendors. Here’s What You Should Steal From That.
By the end of this piece, you’ll know why companies are dumping premium AI models, and whether you should change anything about how you use AI. I read this week’s cost reports, the Stanford AI Index, and the Linux Foundation research so you don’t have to. The verdict up front: open-source AI is now good enough for most work at a fraction of the price, companies figured it out first, and there are three moves you can copy this week for free. Premium AI isn’t dead. But blindly paying premium prices for everything? That era just ended.
Now let me tell you about the CEO who started my week.
The guy who fired the best AI on the market
Flo Crivello runs an AI startup called Lindy. In June he did something that would’ve gotten him laughed out of the room in 2024. He moved 100 percent of his company’s traffic off Anthropic’s Claude models and onto DeepSeek, a Chinese open-weight alternative. His description of the cost curve afterward: it crashed to the ground. He claims the switch will save millions within months.
Quick honesty check, because that’s the whole brand here: that number is his, self-reported, not audited. But the direction is very real, and he’s just the loud version of something happening quietly inside thousands of companies right now.
I know what you’re thinking. “Cool, corporate cost-cutting, why do I care.” Because the reasons they’re switching are the exact same reasons you’re probably overpaying and overexposing your data right now.
Why they’re jumping ship
The performance gap collapsed. Stanford’s 2026 AI Index found the best closed model leads the best open model by just 3.3 percent. Two years ago, that gap was a canyon you could hear echoes in. Z.ai’s GLM 5.2 landed within about a percentage point of Anthropic’s Opus 4.8 on a closely watched agentic benchmark, at roughly a fifth of the cost. One benchmark isn’t gospel, but the pattern repeats everywhere independent people measure.
The bills got scary. Companies spent two years in “adopt AI at any cost” mode. Then the invoices arrived and finance departments discovered feelings. Data from OpenRouter, a platform that routes AI traffic, pegs the new open models at 60 to 90 percent cheaper than leading closed ones. And here’s the secret nobody says out loud: most AI tasks are boring. Summarize this email. Sort this ticket. You don’t need a frontier genius for that. You need a competent free intern.
Lock-in got real. Build your product on one vendor’s API and you inherit their price changes, their rate limits, and their surprise model updates. One quiet change on their end and your product acts different overnight. You find out when your customers do. Fun!
Privacy and sovereignty. Every call to a closed model ships your data to someone else’s servers. Hospitals, banks, and basically all of Europe have opinions about that. The European Commission opened work on a sovereignty-focused open digital strategy in January. Governments now treat open AI as an independence play, not a hobby.
The honest counterweight
Closed models still handle roughly 80 percent of all usage per MIT Sloan research, the hardest problems still belong to the big labs, and enterprises happily pay for someone to call when things break. Openness isn’t guaranteed either: Alibaba just kept its most powerful Qwen models proprietary, reversing its own playbook.
So, this isn’t a funeral for premium AI. It’s a rebalancing. The question changed from “which frontier model do we use for everything” to “what’s the cheapest model that’s good enough for this task.” That shift took eighteen months and it’s permanent.
What you should actually do
Run a local model this week. Download Ollama or LM Studio, both free, grab a small open model, and run it on your laptop. Fifteen minutes. The first time an AI answers you with your wifi off, this entire article clicks into place emotionally. Nothing leaves your machine. Nothing.
Route your sensitive stuff local. Medical questions, financial docs, the business idea you guard like a dragon. A local model never phones home. Best free privacy upgrade available to a normal human right now.
Stop building single-vendor habits. Keep your prompts and workflows in files you own, not trapped inside one AI product. Companies learned this lesson with millions of dollars. You get to learn it for free.
Think like a router. Frontier model for hard reasoning. Free open model for the boring 80 percent. Your subscription budget will send you a thank-you card.
Read the license. “Open weights” is not “open source.” Some models carry restrictions that matter if you build on them. Two minutes of reading, one awkward conversation avoided.
So What
Do this: Install Ollama or LM Studio this week and hand a local model one boring, private task.
Skip this: Canceling your premium AI subscription in a dramatic huff. If it’s earning its keep on hard work, keep it.
Wait on this: Betting your whole workflow on any one open model. The leaderboard reshuffles monthly, so stay portable.
The one-liner to steal: The AI question is no longer “which model is best,” it’s “what’s the cheapest model that’s good enough for this task.”
Companies got the leverage first. Now it’s ours. What’s the first task you’d hand to a free local model?
Sources
Forbes: Open Source AI Is Moving From Sideshow To Strategy (April 19, 2026)
Linux Foundation Research: The Economic and Workforce Impacts of Open Source AI (May 2025)
Axios: Meta to open source versions of its next AI models (April 6, 2026)
TechnologyChecker: Open-Source AI Adoption 2026 (May 29, 2026)