I love having this much intelligence on tap.
I can build something, explore an unfamiliar subject, or follow a question much further than I would have managed on my own. Recently, that meant investigating a Renaissance painting through artists’ relationships, disputed interpretations and Michelangelo’s poetry.
A fairly normal evening, obviously.
The more I use these tools, the more possibilities I see. Then another thought creeps in: being able to pay for this much access is already a privilege.
What happens if that access makes me more capable, helps me create more value, and eventually gives me the money to buy even better access?
Now imagine that happening across companies. Across countries.
How much could that initial advantage compound?
How much thinking can you afford?
“Token maxing” is a slightly ridiculous phrase for something I find quite interesting: pushing these systems to see how much useful work you can get out of them. Tokens are the chunks of text models process and generate; using more of them can mean longer conversations, more attempts and more work happening at once.
Paid AI subscriptions already run into hundreds of dollars a month. Higher tiers can buy substantially more usage, alongside access to particular models or features. Both OpenAI and Anthropic structure their plans around different levels of access.
For someone who uses AI constantly, that expense might be easy to justify. For someone else, it might be completely out of the question.
And the difference can extend beyond the quality of a single answer.
Imagine two people trying to build something. One has enough capacity to test several approaches, investigate why they failed, revise the work and try again. The other has to ration their usage.
Both technically “have access to AI.” Their ability to work with it is quite different.
That freedom to keep going matters. Sometimes the useful insight comes after the obvious question, the unhelpful answer and the follow-up that finally gets somewhere.
I would quite like my curiosity to respect a monthly budget. It has shown limited interest so far.
Useful results still require judgement and some ability to recognise when the machine is confidently talking nonsense. Spending more doesn’t automatically improve either. But the capacity to experiment gives you more chances to learn.
An advantage that could help pay for itself
Wealth has long bought access to help: tutors, advisers, specialist knowledge, people who can do things you cannot.
AI makes me wonder how quickly that relationship could develop.
Suppose better access helps someone learn a useful skill or build a working product. That achievement could lead to income, customers or a better job. Some of those returns could fund more AI usage, alongside more time to experiment.
The advantage begins to accumulate.
There are plenty of ways for that loop to break. A subscription doesn’t automatically produce a successful business. A working prototype still needs customers. A good answer still needs someone to act on it.
Those steps are also where existing privilege matters. Having spare time, useful contacts and enough financial security to attempt something that might fail can make it easier to turn assistance into opportunity.
This is what makes the question uncomfortable for me. I’m excited about these tools partly because I can see what they let me do. I’m also in a position to spend time using them.
Someone else could have just as much curiosity and a better idea, with much less room to pursue it.
What if the best version isn’t for sale?
Then there’s the question of what we can’t access at all.
While I’m busy token-maxing on Astra, feeling quite pleased with how much I can accomplish, I wonder what the next level of access might look like. More computing power? Systems running continuously? Capabilities that haven’t reached a public pricing page?
I don’t know how capable unreleased systems are. Being private doesn’t establish that a model is better, or that any advantage will last. But the possibility is worth considering.
If a company developed a system that gave it a substantial advantage in research, engineering or decision-making, what would make it release that capability to everyone else?
Selling access could be extremely profitable. Keeping some capabilities internal might be valuable too.
What happens if superior AI becomes the company’s moat—the advantage competitors struggle to reproduce?
Even using an available model, an organisation could combine it with its own data, substantial computing resources and people who know how to put the results to work. Matching everything around the model would be a considerable undertaking.
There could be a whole hierarchy of access, with each level looking extraordinary from the one below it.
The same question at the scale of a country
The US and China naturally come to mind here. Stanford’s 2025 AI Index documented their leading roles in producing notable models, alongside a rapidly narrowing performance gap between them. Stanford AI Index
What does this mean for countries with much less money, computing infrastructure or specialist talent?
They can still benefit enormously from models developed elsewhere. Building your own frontier model—a system at the leading edge of AI capability—is an expensive route into AI, and many useful applications don’t require one.
But dependence raises questions.
Who sets the price? Who decides what is available? How well does the technology serve your language, institutions and priorities? What happens if the terms change?
The World Bank describes these gaps through internet access, computing resources, locally relevant data and skills. It also reports substantial AI adoption in middle-income countries, which is a useful reminder that participation already extends well beyond the places building the largest models. World Bank
Still, I wonder how much bargaining power comes from controlling an increasingly important source of productive capability.
If AI becomes deeply involved in how a country researches, builds and runs its industries, decisions about access could have consequences far beyond anyone’s monthly software bill.
The argument against my own argument
There is a substantial reason to be optimistic: useful AI has become dramatically cheaper.
Stanford’s 2025 AI Index reported that the cost of using a system performing at roughly GPT-3.5 level fell by more than 280 times between November 2022 and October 2024. That measures the cost of using models at a particular performance level; it doesn’t mean everyone’s subscription became 280 times cheaper. Still, it is an extraordinary change. Stanford AI Index
Cheap and free tools could give people access to help they previously couldn’t obtain at any price they could afford.
A student could get repeated explanations of something they don’t understand. A small business could attempt work it would otherwise have to outsource. Someone with an idea could reach a first prototype without first assembling an entire team.
Those are meaningful gains.
Models that can be downloaded and adapted offer another route. They still require resources and expertise to run, but they create possibilities beyond relying entirely on a single provider’s hosted service. The World Bank identifies open technologies and smaller, affordable systems as opportunities for developing countries. World Bank
I want that future. It is part of why I’m so enthusiastic about AI in the first place.
Cheaper intelligence. Still behind?
But I’m not sure falling prices answer the question that bothers me.
Suppose everyone eventually gets access to something as capable as the best model I’m using today. By then, what will the people who can afford the newest, most capable systems be using?
Yesterday’s expensive intelligence could become widely available while the people at the top keep moving.
And any advantage they gain along the way might stick.
If better AI helps a company develop a product sooner, win customers and reinvest the returns, making that same capability cheap a year later doesn’t rewind the year. The customers, revenue and experience have already accumulated.
Everyone getting faster doesn’t necessarily mean everyone catching up.
Helping someone learn, build something or solve a problem they previously couldn’t is a meaningful improvement in their life. I wouldn’t dismiss that. But when we talk about competition—who wins a contract, builds the leading company, owns the infrastructure or gains influence—the relative advantage matters too.
That leaves one crucial question: how much better does the most expensive, least accessible intelligence remain?
If affordable models become good enough that further improvements make little practical difference, the gap could lose its importance. Knowing your field, choosing a useful problem and following through might matter far more than your subscription tier. That is the strongest version of the optimistic argument.
If the best systems continue to offer a substantial advantage, though, falling prices alone won’t reassure me. The people who can afford them could keep turning their lead into the resources needed to maintain it.
Universal access to useful intelligence would be an enormous achievement. I still want to know whether it gives everyone a realistic chance to catch up.
I’ll keep using these tools. I’m far too interested in what they make possible to pretend otherwise.
But every time I get excited about how much further I can take an idea, there’s another question worth asking:
Who else gets the chance to keep going—and who gets to stay ahead?