If you’ve been using AI consistently over the past year, you’ve probably started to feel a shift that’s hard to ignore. It’s not always obvious at first, but over time it becomes clear—usage limits are tightening, responses feel slightly more constrained, and you’re getting less output for the same amount of effort. What used to feel fluid and almost limitless now feels a bit more controlled, a bit more metered, and in many cases, a lot more expensive.
This isn’t just a perception. Across multiple platforms, there has been a noticeable tightening of access and value, even on paid plans. Tools that once felt incredibly generous with usage are now introducing stricter caps, more aggressive throttling, and clearer segmentation between what you get at each pricing tier. You can see this happening with providers like Claude and Cursor, where the experience has gradually shifted from open-ended exploration to something that feels more like a utility with a meter running in the background.
And it’s not just about pricing, which is what makes this even more interesting.
It’s not just getting more expensive, it’s getting less capable at lower tiers
There’s another layer to this shift that doesn’t get discussed enough, and that’s the perceived drop in capability depending on the tier you’re using. The free versions and even some of the lower-tier paid plans are starting to feel noticeably less powerful than they once did. The responses are often more surface-level, the reasoning feels less robust, and tasks that previously required minimal prompting now take multiple iterations to get right.
In many cases, these models are starting to feel more like advanced autocomplete systems rather than true reasoning tools. You can still extract value, but it requires more effort, more structured prompts, and more back-and-forth to reach the same level of output that once came more easily. Naturally, this leads to more usage, more tokens being consumed, and ultimately higher costs for the same end result.
The era of having access to top-tier AI capabilities for a flat monthly fee is starting to fade. What we’re moving toward is a much more usage-driven model, where the depth of interaction and the complexity of your tasks directly impact how much you end up paying. And for teams that rely heavily on AI, that shift has real implications.
The idea of fully automated AI teams starts to break down
Not too long ago, there was a lot of excitement around the idea of building fully automated teams powered by AI. The vision was compelling—design, engineering, support, and operations all orchestrated through a combination of LLMs and automation tools, running continuously in the background and dramatically reducing the need for human intervention.
And to some extent, that vision is still valid.
However, the economics of making it work at scale are becoming harder to ignore. When you start chaining together multiple AI calls, especially those involving reasoning-heavy or “thinking” models, the cost begins to add up very quickly. Workflows that rely on iteration, where models refine their own outputs over several passes, can burn through a significant number of tokens before arriving at something usable.
This becomes even more pronounced when you’re working with automation tools that allow models to operate with a degree of autonomy. If those systems are not tightly controlled, they can end up consuming large amounts of resources in the process of “figuring things out,” often without delivering proportional value.
What initially felt like an efficient, scalable approach starts to look expensive and, in some cases, inefficient. The idea of letting AI run freely until it reaches the right answer becomes less appealing when every iteration has a cost attached to it.
The reality is that AI was never truly cheap
The more you look at it, the clearer it becomes that AI was never actually cheap to run. For a long time, the cost was simply hidden or absorbed by the companies providing these services. Generous free tiers and relatively low subscription prices made it feel accessible, but that accessibility was largely subsidised.
Now we are starting to see what happens when that subsidy begins to disappear.
As companies look for sustainable business models, the real cost of running large-scale AI systems is being passed down to users in a much more direct way. At the same time, reliance on these tools has increased significantly. Many individuals and teams have built AI into their daily workflows to the point where it is no longer optional—it is a core part of how they operate.
This creates a situation where users are both more dependent on AI and more exposed to its true cost. From a business perspective, it’s a logical moment to adjust pricing. From a user perspective, it feels like the ground is shifting beneath your feet.
So what should teams actually do?
Despite all of this, AI is not going anywhere, and it remains incredibly valuable when used correctly. It still has the ability to improve productivity, accelerate development, and unlock new ways of working that were not possible before. The key shift is not whether to use AI, but how to use it more intentionally.
Instead of treating AI as an unlimited resource, teams need to start thinking about efficiency. This means being more deliberate with prompts, reducing unnecessary iteration, and designing workflows that minimise waste. The clearer your thinking is before you engage with the model, the less you rely on expensive trial-and-error cycles to get to a good result.
AI works best when it is augmenting well-defined processes, not when it is being used to compensate for a lack of clarity.
The growing shift toward local models
One of the more interesting developments in response to rising costs is the increasing interest in running models locally. What once felt like something only accessible to large organisations or highly technical individuals is becoming more achievable for smaller teams and even solo builders.
Advances in hardware, combined with improvements in model efficiency, have made it possible to run certain types of models on consumer-grade machines. Devices like the Mac mini, especially when configured with sufficient memory and connected through high-speed interfaces, are becoming a practical entry point for local experimentation.
This approach offers a different set of trade-offs. While local models may not yet match the absolute performance of the latest cloud-based systems, they provide greater control over cost and data. For teams that are concerned about long-term expenses or the privacy of their intellectual property, this becomes an attractive option.
There is also a growing ecosystem of tools and frameworks designed to support local deployment, which suggests that this is not just a niche trend but a direction that will continue to develop.
Where this leaves us
AI is still one of the most powerful tools available to product teams today, but the way we interact with it is changing. It is no longer something that can be treated as infinite, and it is no longer priced in a way that encourages careless usage.
In many ways, this shift is forcing a return to fundamentals. Clear thinking, well-defined problems, and structured workflows are becoming more important again, because they directly impact how efficiently you can use AI.
The tools are evolving rapidly, but the principles of building good products remain the same.
If you’re a founder or part of a product team trying to navigate this shift—figuring out how to use AI effectively without letting costs spiral—subscribe to the Dive Into Product blog.
I’m sharing how we’re adapting our workflows, where AI is delivering real value, and where it’s starting to fall short.