Your AI Is Confirming What You Already Think. That's Not Even the Real Problem

Your favorite AI companies are going public and you heard prices may go up. Meanwhile, your AI is mostly giving you beautiful flattery. What do we do?

There's always been a lot of talk about the sycophancy problem with AI. The tendency of these tools to agree with everything you say, prop up your dumb ideas and confirm what you already think is a challenge. The proposed fix is usually to set your AI up to push back, challenge your assumptions and argue with you. Unfortunately, that's the same problem turned upside down.

Luckily, the solution is pretty simple, although it is also hard. The only important question we have to ask is this: Is AI making you more of who you are trying to be or is it not?

The Context Problem

If you already know what you're trying to expand in yourself, in how you think, in how you participate in the world and in what you're building, then AI is just another tool in service of that. It can now do things that you didn't explicitly direct, surface things you didn't ask for, and help test pathways you would not have had time to test alone. The issue is whether the system has enough context to serve the work without flattening it.

Without that anchor, AI amplifies noise and makes the existing pattern more efficient, whether or not that pattern deserves to scale.

What helps here is that the destination and the pathway are two separate questions. You don't have to know exactly how to get somewhere in order to start. There are always multiple pathways to the same place. AI is particularly good at helping you test them, collect data on what's working, and iterate faster than you could on your own. If the work has a real center but the path is still emerging, that's exactly the kind of problem AI can help you work through.

All of this needs to be coupled with the practical stuff: Knowing which model to use, understanding what you're paying for AI with, understanding why it might be important for you to learn AI or enlist the help of someone who has, and having a grip on how major shifts in the AI economy affect your future AI use.

What You're Actually Paying For

The other day, I tested Anthropic's Fable against Anthropic's Sonnet (before Fable got suspended). Fable is Anthropic's newest model and the assumption out there is always that newer means better. I wanted to run tests because I never actually found a good use for Opus, their previous top tier. The outputs always felt slightly off. It seemed like the reasoning went one step past reality and landed somewhere adjacent to what I actually needed. So, after realizing that Opus was way overpowered such that it was overthinking all my writing, code and content, and was actually making my life harder, I've been running almost everything on Sonnet ever since. That saved me a lot.

But, I get interested in what's possible too since I have been astounded by how the world of possibility seems to have cracked open this year. Anthropic has been talking about Fable for awhile because there has been all this hype about another one of their models called Mythos. Mythos was said to present severe security risks to the world. Fable is just Mythos with its teeth removed. Before Fable got pulled back after only being out and about for a few days, it had a variety of guardrails built in that were meant to steer users away from taking harmful action with the model.

So I ran some tests. Fable was noticeably more elegant in its solutions. I'd say 40-50% more, if I had to put a number on it. It's also three times the price of Sonnet, didn't get to the results any faster, and required about the same amount of back-and-forth to get there. So what you're paying for with Fable is more elegant output and solutions, at three times the cost, for the same amount of your time, assuming you are doing relatively normal people activities. The tests I had it run were around fixing media creation workflows, like one that pulls down clips from Youtube and edits in AI generated images to create reels for Facebook. The video from that test did great on Facebook by the way. It was about Stoicism versus the current political climate.

The one more useful test I ran with Fable was having it plan new features for my agents. It was a joy to use it in that way.

If you are running extremely complex tasks, your results with Fable will probably be more impressive than mine. But a lot of useful business AI is not exotic. It is non-thinking work: video editing, media creation, spreadsheets, writing, research, workflows that previously required three different softwares. All of this can be done well by cheaper models like Sonnet and often doesn't get done well by more advanced models because the model overthinks and overcomplicates the task.

For writing, analysis, basic coding, research (most of the work people actually do) over-reasoning is a liability. You don't want your AI to have thought about something so hard that it guides you somewhere wrong and charges you extra for it.

What You're Actually Paying With

All of this is important to know because there's a lot of talk about how AI companies going public will inevitably lead to prices skyrocketing. The argument is that as these companies move toward IPO and training costs stay extremely high, prices have to go up. Which may be true. But before reaching that conclusion, it's important to distinguish what you're actually paying for and what you're paying with.

There are two primary ways people are paying for AI. Consumer subscriptions, which are anywhere from $20 to $200 a month, and API access. These are not the same product, and the difference isn't just price.

Consumer subscription plans have lower data privacy protections. Anthropic updated their consumer terms in August 2025: conversations from Free, Pro, and Max subscribers are used to train their models unless you actively opt out. Data can be retained for up to five years. The change went out as an in-app notification with a settings deadline. It did not make headlines the way it probably should have.

API access is different. At Anthropic, API data is retained for 7 days for abuse monitoring and never used for training. At OpenAI, it's kept for 30 days, and also not used for training. Enterprise and Team accounts across both are excluded from training use by default. This is why enterprise companies often end up on API. It's a legal requirement, not a preference.

What this leads us to understand is that data privacy is the real commodity here, not AI access. The subscription model is the same structure as Facebook: access is subsidized because your conversations are a resource these companies need. We've been conditioned to hand this over without thinking much about it, in exchange for tools that feel low-cost.

Which is also why subscription prices are harder to raise than the narrative suggests. These companies are getting consumer training data, a key resource they need to continue developing, at essentially no cost. Raising prices significantly risks losing the pipeline that makes training new models cheaper. The ceiling on what can be theoretically developed with AI is very high and we are nowhere near it.

Why AI Prices Probably Won't Spike

AI companies already do use synthetic data to train their models but as we all know, it does not take long for AI to abstract itself into oblivion without human anchors. And, if synthetic data was as useful as some might like to believe, companies like Facebook would likely have already moved in that direction, but they haven't. Facebook (and social media in general) remains a copious consumer of user data. There aren't a lot of reasons to believe that AI will be any different. The other important development that we will see in the next year is around AI getting metaphorical eyes and ears as it is deployed in settings that aren't computers, like wearables, robots, etc. All of that will have to be trained and the only way to have that happen (so that people actually adopt these tools) is to subsidize it.

Enterprise plans and API aren't going to be easy to raise prices on either. The only way to get real live companies to be willing to go through the growing pains of adopting AI into their workflows is to subsidize it. There is a lot of institutional and regulatory resistance to incorporating AI. Everyone knows AI can do cool things but most companies aren't willing to pick up a tool just because it's theoretically cool if it's going to also cause a lot of pain, unless it is subsidized. Without those subsidies, many companies are not going to want to go through the AI adoption process at the rate that the AI providers need them to.

Another counterweight to the argument that prices will definitely go up after IPO relates to Chinese models. DeepSeek and others are orders of magnitude cheaper for outputs that are comparable to top-tier American models. The concern, depending on how you think about the world, ranges from the paranoid (although that may be fair): the Chinese government has access to data on large numbers of Americans at scale. What does that mean? At the other end of the spectrum is a strategic concern: By using Chinese models, you're funding Chinese AI development over American AI development. I think both are probably important considerations.

On the other hand, many of these Chinese models are becoming available to run locally, on your own computer, in which case your data does not leave the device and is therefore not getting processed on a server in China. There are of course concerns that there could be backdoors to these models that cause data security issues, but that is not stopping many people from starting to go that route. This puts further pressure on American companies to keep their pricing accessible.

Open source tooling is the other piece. There's an ecosystem of free tools that make AI agents work more efficiently, resulting in lower token use, better results and less back and forth. These tools allow users to make their AI quotas go further, which would be somewhat protective if prices did go up.

What to Do Now

The fear of prices going up in the short to medium term is probably unfounded. After another 3-5 years, we might be having a different conversation. That means, now is the time to use and learn AI, since costs will probably stay low for a bit. If there are projects, visions, or bodies of work you have not had the capacity to build yet, now really is the time. Don't wait. The toward-what question is yours to answer. These conditions being favorable just means the window is open. What you build in it and whether or not it actually expands your capacity to think and to be present with people, is a separate question that would be good for each of us to answer as soon as possible.

If you don't want your data used to train models (and it is a very big open question of what might happen in this conversation if an event caused everyone to pull their data out of training all at once since that would eliminate one of the reasons we will likely keep seeing subsidization of AI), you should opt out in whatever provider you are using.

It also means that, while prices might not be changing soon, it would be wise to know your alternatives. These alternatives will likely be running models locally and using open source tools to keep your AI running as efficiently as possible. You don't need to become a computer wizard in order to start developing an understanding of what's involved in that process, but you do need to move out of the chatbot world and into the agentic world, at least, to begin developing your understanding.

The sycophancy problem is just a symptom of the broader issue. AI that confirms what you already think or executes what you already do, just faster, is clearly not good enough. But the fix isn't an AI that argues back for its own sake either. It's building enough context around the work that AI can serve the thing you're actually trying to expand and you can notice when something is pulling you away from it.