If tokens are the new drug... are Anthropic and OpenAI the new cartels? 🤔 Get 'em hooked for free and then start charging. It's the oldest trick in the book. And the AI giants have been running this for a while. Your Pro/Max subscription is heavily subsidized by hundreds of billions of $$$ of investors' money. A $200 Max subscription can easily consume thousands of dollars worth of tokens per month. It's not about making money, it's about making us addicts (and I'm only half joking here). And I say this as a full-blown addict. I've built skills, agents, automations - a whole personal operating system running on #Claude. The dopamine discharge when something you built in 5 minutes just works? Out of this world! It's no longer a SaaS subscription, it's a habit! None of this is new of course, but here's where it gets interesting. Last week Anthropic announced that starting next month, your subsidized subscription stops covering programmatic use: the SDK, claude -p, the automated workflows. Exactly the stuff the heavy builders (like me) built their lives on. Building? Still subsidized. Running the agents you can't live without? Start paying! But (and this is the best part) they're smart enough not to cut you off cold turkey, so you don't go shopping for a new dealer (hmm... OpenAI?). Instead, they make a gesture and give you credits - $200 for a Max 20x account. Enough to make you feel fine at first, so you keep building, keep depending, keep getting deeper. Until one day the credits are just not enough, and there's no path back. That's not a bug, and it was the obvious move from the get go. That's the business model working exactly as designed. But it's still worth pointing out - especially by someone like me who's completely sold on how transformational this is. Both things are true: the technology is genuinely remarkable, yet the people building it know precisely how good a dealer's hook feels. So this begs the question - are LLMs the most productive tool of our generation, or the most elegant dependency ever engineered? (Or both.) Asking "for a friend." Who may need a sponsor. 😅
I think the real lock-in is not the model itself. It’s the accumulated workflows, architecture, memory, orchestration, and operational behavior systems built around it over time.
Nailed it, Ofer. We've run this movie before. VCs spent a decade subsidizing our Uber rides and $4 DoorDash burritos to rewire how a whole generation moves and eats. The subsidy ended, the real price showed up, and by then the habit was load-bearing. Tokens are the same play, just faster and stickier, because this time the habit lives inside how we build. The part that nags at me sits one layer deeper. The dependency you're describing runs to one provider's pricing mood. The builders who stay free can switch dealers without withdrawal. They know what each task is worth and route it to whoever delivers it best per dollar. So maybe the sharper question than "productive tool or elegant dependency" is this: did you build on one model, or on optionality? (And happy to sponsor you. I'm a few skills deep myself. 😅)
Now imagine youre in chemistry or biology discovery with these models. You’re paying for these tokens/compute + all the downstream lab tasks that need to be done. There maybe no way around these models anymore in the lab.
The dependency angle is real, but I think the stronger framing isn’t “cartel,” it’s infrastructure pricing catching up to actual compute cost. What’s changing now is that the free/cheap layer was always a land grab for workflows — once those workflows become critical, usage-based pricing is inevitable, especially for agentic + programmatic loads. The interesting risk isn’t addiction, it’s hidden coupling: when your “personal OS” is tightly bound to one provider’s runtime assumptions. Ofer Maor
I hope token costs will be reduced to nothing over time
Anthropic and OpenAI are powerful AI platforms, but framing them as “cartels” is rhetorical; real issue is subscription dependency and usage-based cost scaling.
The dopamine discharge is real great post
First taste is always free!
I have paid the 20$ subscription on Claude code. It is true that it is really good, you receive high value for that price. On the other hand, I am also testing open source models and planning to switch in the future to complete open source. Open source is the way to go for me as a developer, not planning to give too much power to a single company like anthropic.