TL;DR
- Cost optimization matters, but doing it too early limits what your product can do
- OpenClaw’s approach: don’t cut corners, don’t downgrade — let users bring their own AI subscriptions
- Three product shifts for the age of abundant intelligence: Outcome over Output, Brute Force Wins, Disposable UI
- Before your product proves it deserves to exist, the biggest risk isn’t spending too much — it’s building something nobody cares about
What Is Abundant Intelligence?
Inference costs are dropping by orders of magnitude every year. The price wars between model providers keep getting fiercer. A top-tier model that felt too expensive to use a year ago now costs a tenth of what it did. This trend isn’t slowing down.
When intelligence stops being a scarce resource, everything shifts — how software works, how products are designed, how business models function. Most of us, myself included, haven’t caught up yet.
My First Instinct When Building AI Products: Cut Costs
When I’m planning the architecture for an AI product, my brain immediately goes to cost. Can I swap OpenAI’s API for an open-source model? Can I shrink the context window a bit more? If users interact this way, what’s the per-user cost?
I keep asking myself: can this scale?
But here’s what I noticed. The AI products that actually feel great to use? They’re almost all burning tokens without hesitation. When I use Claude Code, I never think about how many tokens I’m using. I just focus on solving the problem. And when it helps me get things done, I’m happy to pay.
Saving Tokens Isn’t Wrong — But the Timing Is
Cost optimization is fine. Every mature product eventually needs to deal with unit economics.
But a lot of founders — especially technical ones — start making technology choices based on cost before they’ve even found product-market fit . They pick models by checking the pricing page first. They design features around token budgets. They deliberately limit what their product can do because it’s “too expensive.”
It’s like a restaurant researching cheaper ingredients before they’ve figured out if the menu is any good.
The token costs you spend months of engineering effort optimizing today? A single model upgrade six months from now could make all that work irrelevant. Spending your scarcest startup resources — time and attention — on a problem that’s rapidly depreciating in value is missing the forest for the trees.
OpenClaw: A Counterintuitive Case for Brute Force
Peter Steinberger, the creator of OpenClaw, took an approach that goes against the mainstream: don’t cut, don’t shrink, don’t downgrade. (For more on Peter and the OpenClaw story, see my earlier post: When AI Learns to Rifle Through Your Drawers)
OpenClaw connects through ChatGPT Codex via OAuth, letting users drive the agent with their own subscriptions. Developers don’t need to absorb API costs, and users — already paying a monthly fee — don’t stress over every agent call.
Peter’s logic is straightforward: if an agent can save an engineer 2 hours of debugging, even if that run costs $5 to $10 in tokens, it’s a no-brainer. The real question was never “how much does this API call cost?” It’s “is the model smart enough to get the job done?”
He calls this brute force . Sounds rough. But think about it — when you stop worrying about every token, you finally have room to focus on what actually matters: how big of a problem can this product solve for users?
The Cost Dilemma: Who Pays the Bill?
There’s a real structural problem here.
Most AI products aren’t stuck with weak models because they want to be. They’re stuck because they can’t afford the good ones. When you’re calling models through APIs, every user action shows up on your bill. More users, more pressure. That structure naturally pushes developers toward cutting costs.
OpenClaw offers a different model: let users connect their existing AI subscriptions — ChatGPT Plus, Claude Pro — so developers don’t have to run every feature decision through a cost review.
This immediately raises another question: if the core compute comes from users’ own subscriptions, where’s the moat? Big question, different article. But for now, at least in the early exploration phase, this model frees you from cost anxiety so you can focus on building something people actually want to use.
Three Product Shifts for the Age of Abundant Intelligence
When tokens are no longer scarce, product design looks very different.
Outcome over Output
Users want their problem solved, not a chat interface.
With abundant intelligence, a system can run dozens or even hundreds of self-correction cycles in the background — burning through compute just to make sure the final result is correct and actionable. Users don’t need to know how many rounds it took. They just need to see: it’s done, and it’s right.
There’s an economics concept called Jevons Paradox: when a resource gets cheaper, total consumption goes up. Same with AI — the cheaper intelligence gets, the more “wastefully” we’ll use it. But that waste is exactly what users experience as better results.
Brute Force Wins
An extension of the last point. If the problem is valuable enough, throw enough resources at it.
An agent that saves a user an entire afternoon is worth paying for, even if a single run costs several dollars. An AI assistant that can read through ten years of a user’s journal entries and generate a deep personal profile — that experience versus “20 messages per conversation” is the difference between a product people love and one they forget.
Disposable UI
If intelligence is powerful enough, fixed software interfaces become redundant.
An agent can generate a form, a chart, a set of buttons on the fly based on the current task — then discard them when it’s done. Developers no longer need to pre-design comprehensive interfaces, because every interaction gets a UI built specifically for that moment.
This isn’t science fiction. When inference costs are low enough to ignore, generating a fresh UI every time isn’t waste — it’s the most sensible approach.
If You’re Calculating Your Token Bill Right Now
If you’re building an AI product and spending more than 30% of your mental energy figuring out how to save on API costs, step back. Is overspending really your biggest risk at this stage?
More likely: you build something that works but nobody cares about. Users try it once and leave. The experience isn’t good enough. And a lot of the time, the experience isn’t good enough because you didn’t let the model do its job.
The age of abundant intelligence is coming. Let your product prove it deserves to exist first. The bill can wait.