A few days ago, Anthropic quietly removed Claude Code from their Pro plan and made it exclusive to their Max tier. Then they walked it back after backlash. Around the same time, Microsoft started tightening Copilot access, limiting new sign-ups and imposing stricter usage caps on existing users.
If you've been building workflows or products on top of these tools, you probably felt a small jolt of unease. You should. This isn't a blip. It's the beginning of a pattern we've seen play out in every VC-backed platform that captured a market by running services at unsustainable prices.
The question isn't whether AI pricing will change. It's how fast, and what you're going to do about it.
The Claim or Trend
AI companies are starting to pivot from growth-at-all-costs to profitability. The cheap, powerful access we've enjoyed over the past 18 months is beginning to tighten. Usage caps are appearing. Features are being gated behind higher-priced tiers. Some companies are testing how much they can restrict without losing too many customers.
This follows the exact playbook we saw with Uber, Deliveroo, and Netflix. Capture the market with artificially low prices funded by venture capital, build dependency, then shift the economics once you've got critical mass.
For AI, the constraints are real. Semiconductor shortages, data centre capacity limits, and the sheer cost of running inference at scale mean these companies can't keep burning money indefinitely. The pivot was always coming. It's just arriving faster than most people expected.
Why It Sounds Compelling (And Why It's Uncomfortable)
The narrative around AI tools has been relentlessly optimistic. You can do more with less. You can handle bigger workloads. You can replace entire functions with a well-crafted prompt and a decent API budget.
For a lot of developers and agencies, this has been true. I've spoken to people who've genuinely taken on more clients because they can automate research, drafting, and even parts of code review. The productivity boost is real.
But that productivity is built on a foundation that was never designed to be permanent. It was designed to be cheap enough to get you hooked.
Now the economics are shifting, and the people who built their workflows entirely around these tools are starting to realise they don't control the infrastructure. They don't control the pricing. They don't even control whether the features they rely on will still exist next quarter.
That's not a comfortable position to be in.
What Actually Holds Up
The core capability of these models isn't going away. Large language models are genuinely useful. They're good at summarising, drafting, translating context between formats, and handling repetitive cognitive tasks that don't require deep domain expertise.
The problem isn't the technology. It's the delivery model.
When you're running workflows in n8n that call out to Claude or GPT-4 dozens of times a day, you're not just using a tool. You're depending on someone else's infrastructure, someone else's pricing strategy, and someone else's capacity constraints.
If that pricing doubles, your margin shrinks. If they introduce rate limits, your workflow breaks. If they decide to gate a feature behind a higher tier, you either pay up or rebuild.
This isn't hypothetical. It's already happening.
Where This Breaks Down
The assumption that drove a lot of early AI adoption was that prices would stay flat or even decrease as models got more efficient. That was never realistic.
Yes, models are getting more efficient. But demand is growing faster than efficiency gains. Data centre capacity isn't infinite. Semiconductor supply chains are still constrained. And most importantly, these companies are under pressure to show a path to profitability.
The result is a pricing environment that's going to get more expensive, not less. Features that were bundled into base plans are going to move to premium tiers. Usage caps that didn't exist are going to appear. And the companies that can't make the economics work are going to either raise prices aggressively or shut down.
Google might be the exception. They've got the infrastructure, the capital, and the strategic patience to run AI services at a loss longer than most competitors. But even Google isn't immune to the same constraints. They're just better positioned to absorb them.
For everyone else, the pivot is coming fast.
Implications for Automation and n8n
If you're building automation workflows that depend heavily on external AI APIs, you need to start thinking about resilience.
That doesn't mean abandoning cloud AI entirely. It means not building your entire operation on a single vendor's pricing model that could change next quarter.
One approach that's starting to make a lot of sense is hybrid architecture. You run n8n locally or self-hosted, and you use a cloud database like Neon or Supabase as your backend. This gives you the flexibility to keep your workflows running even if external services change their terms or pricing.
You're not locked into a SaaS platform's infrastructure. You're not dependent on their uptime or their rate limits. And if you need to swap out one AI provider for another, you can do it without rebuilding your entire system.
This is especially relevant for agencies or teams that are running client workflows. If your client's automation breaks because Anthropic changed their pricing, that's your problem. If you've built redundancy into the system, you've got options.
How I'd Approach This in Practice
The shift I'm seeing, and the one I think makes the most sense, is towards local AI where feasible.
Running models locally isn't realistic for everything. If you need GPT-4 level reasoning or you're processing huge volumes of text, you're still going to need cloud APIs. But for a lot of tasks, smaller models running on decent hardware are good enough.
The barrier to entry for local AI is dropping fast. Computers are getting more powerful. Open-source models are getting better. And the tooling around running and fine-tuning models locally is improving.
If you're using n8n, you can build workflows that call local models for simpler tasks and fall back to cloud APIs for the complex stuff. That gives you cost control, resilience, and flexibility.
It also means you're not completely at the mercy of pricing changes. If Claude doubles their API costs tomorrow, you don't have to absorb that across your entire operation. You've already offloaded the bulk of your usage to local infrastructure.
This isn't about being paranoid. It's about being realistic. If your business depends on AI tools, you need to control as much of the stack as you can.
The Metadata Leak and What It Tells Us
There's a small but telling detail in all this. Cursor's composure model, which a lot of developers have been using, is apparently built on Kimmy. We know this because Cursor forgot to strip the metadata from one of their commits.
This isn't a scandal. It's just a reminder that even the tools that feel polished and proprietary are often built on top of other models, other infrastructure, and other dependencies.
The more layers of abstraction you add, the more points of failure you introduce. And the more you depend on someone else's stack, the less control you have when things change.
Local AI isn't just about cost. It's about reducing the number of external dependencies that can break your workflows.
What This Means for Developers
If you're not using AI tools yet, this might actually be good news. The hype cycle is cooling. The pricing is stabilising. And the people who went all-in on AI without thinking about sustainability are starting to hit constraints.
That doesn't mean AI isn't useful. It means the market is maturing, and the people who survive are the ones who built resilient systems, not the ones who assumed cheap access would last forever.
If you are using AI tools heavily, now's the time to audit your dependencies. How much of your workflow breaks if one provider changes their pricing? How much of your margin disappears if API costs double? How much of your productivity depends on features that could be gated behind a higher tier next month?
These aren't comfortable questions, but they're necessary ones.
Closing Thoughts
The AI pricing pivot isn't a surprise. It's the natural evolution of any VC-backed platform that captured a market by running services below cost.
The surprise is how quickly it's happening, and how many people built their workflows without planning for it.
If you're using AI tools in production, you need to start thinking about resilience. That means hybrid infrastructure. That means local models where feasible. That means not pegging your entire operation to a single vendor's pricing model.
The tools are still useful. The technology is still powerful. But the economics are changing, and the people who adapt early are going to have a much easier time than the ones who wait until their workflows break.
This isn't about abandoning AI. It's about using it sustainably, with redundancy built in, so you're not completely at the mercy of someone else's pivot.
