The terms nobody is talking about
Claude Code updated its usage policies on April 4th. Claude subscriptions are now only available for use in Claude Code. A #Claudepocalypse for individual users (especially if you were using your Claude Code with OpenClaw), but fine — Anthropic gets to set its own terms.
But then came the second part — the part that is being quietly rolled out, first to the biggest Claude customers….
Businesses can't buy Claude subscriptions at all. The only option for any enterprise organization that wants to use Claude officially will be the API.
And while that sounds shocking on the surface, it shouldn't be. The idea that AI could be priced like a gym membership — a fixed monthly cost for effectively unlimited usage — was always an illusion. We've written about this before, probably more times than anyone wanted to hear, but the underlying point hasn't changed: inference is infrastructure, and infrastructure has always been priced based on consumption. Flat-fee inference forces providers to "manage" costs, which inevitably shows up as a worse user experience — rate limiting, shrinking context windows, and downgraded models.
What's actually changing right now isn't the economics, it's the removal of the subsidy. For a period of time, products like Claude Code absorbed levels of usage that would otherwise cost thousands of dollars a month. One developer tracked it precisely: 10 billion tokens across eight months on a $100/month Max plan – $15,000 in API-equivalent value, $800 paid. That worked as a growth lever, but as Anthropic reportedly gears up for an IPO soon, those subsidies were always going to disappear.
The age of practically free inference is coming to an end.
What happens next: a play in three acts
So what happens next? Will engineers go back to coding by hand?
Unlikely. But things will probably get weird for a bit.
Act one: Claude Code usage goes underground
It starts with one developer expensing a personal subscription. Then an engineering manager puts five personal subscriptions on a credit card. Then the org just starts handing out credit cards so every engineer can expense their own personal Claude subscription.
Shadow IT. Individual accounts. No centralized access control. No security posture. No audit trail. In the course of writing this article, we've spoken to no fewer than five developers who admitted to using personal Claude Code subscriptions at work.
And of course this will kick off a new chapter of the cat and mouse enforcement game, with Anthropic hunting and shutting down subscriptions that look like commercial use.
This isn't an attempt to cut corners, it's trying to comply with a fast-arriving mandate. API pricing isn't the problem — rug pulling is. It's like telling everyone with a bus pass they now have to commute by taxi next week.
When the CISO asks how you lost visibility into which AI tools your engineers are using, the honest answer will be: Claude Code's Terms of Service made doing this responsibly impossible.
Act two: pay API pricing
Over time, most companies will do their best to get in compliance with the new normal. They want to do this the right way — dealing with legal, finance, and leadership scrutiny while trying to justify spend and put real governance in place. They're not trying to go rogue.
As companies adjust to the new normal, they'll shift to usage-based pricing (again, just like all other infrastructure products).
But once you're operating this way, a new question arises: why limit yourself to only one model family? There are many models with excellent performance at a fraction of the cost of frontier models from providers like Arcee AI, BytePlus, MiniMax, Moonshot AI, Zai, and others, often at very different price and performance points depending on the use case.
That's the real shift API pricing drives. You're no longer choosing a single tool and living inside its constraints. You're making decisions about how to route, combine, and optimize inference across an evolving ecosystem. And once you're paying by the token anyway, there's very little reason to stay locked into a single family of models — or a single provider.
Act three: an open model future
At some point, this stops being about Claude Code entirely. Once teams are operating on usage-based pricing, the center of gravity shifts. You're no longer choosing a single model or a single interface — you're deciding how to run inference across an ecosystem.
That shift leads somewhere pretty quickly: an open model future.
The best model changes. The cheapest model changes. New models show up, others fall behind, pricing moves, capabilities improve. Locking yourself into a single provider stops making sense in a world that's evolving this quickly. So instead of standardizing on a model, companies start standardizing on how they build, while staying flexible on what runs underneath.
That's the world we built Kilo for.
You can use your Anthropic API key, or tap into 500+ models through the Kilo Gateway — choosing the best option for each task.
With Kilo, you get the agentic experiences that matter for your company — from developer workflows to always-on agents with KiloClaw. You can standardize workflows across your team while routing across models, adapting as the landscape evolves without having to rebuild everything every time a provider changes pricing or policy.
The difference isn't just model access or flexibility. It's that Kilo is built for how companies actually use AI.
The subsidies are ending
Anthropic deserves real credit for bringing agentic coding into the mainstream. Claude Code showed what was possible and helped accelerate adoption across the industry. But that success made the economics impossible to ignore. The generous subscription pricing — tools worth thousands for $200 a month — was never going to last forever for everyone.
And you can already see how this is playing out. Claude Code is just working through the list in order. Enterprise first, because that's where the money is. Mid-market and SMB are probably not far behind.
If you haven't gotten the call yet, that's not a signal that you're safe. It just means you're not at the top of the list. The underlying shift is already in motion, and it applies broadly.
The era where individuals and businesses could both benefit from the same subsidized pricing model is coming to an end. But that doesn't mean the situation is getting worse — it means it's normalizing. Inference is becoming what it was always going to become: infrastructure.
And once you look at it through that lens, the implications are fairly straightforward. You use the models that perform best for your use case, you pay based on actual usage, and you avoid tying yourself too tightly to any one provider's pricing or product decisions. In that world, model-agnostic tools aren't just helpful — they're how you stay in control as everything else continues to change.
Top comments (0)