Jacob Whitmore

10–15 minutes

The Intelligence Divide: What Happens When AI Becomes Too Expensive for Ordinary People?

Checked 31 October 2026. Basic AI is collapsing in price while the frontier gets dearer to run. That gap is the Intelligence Divide, and it decides who benefits most.

By Jacob Whitmore

Checked 31 October 2026

AI was supposed to be for everyone

On 30 October 2026, people paying OpenAI $200 a month for ChatGPT Pro lost half their Work and Codex usage. The price stayed the same. A month earlier, at DevDay, OpenAI had launched a $500-a-month plan, the only one with its fastest “Ultrafast” mode (TNW, 29 Sep 2026).

Nobody marched over this. Most people will never pay $200 a month for software. But it runs against the story we were sold.

That story said AI would democratise intelligence. A student with a cheap laptop would get the same research assistant as a partner at a law firm. A one-person business would get a developer, a copywriter and an analyst for the price of a streaming subscription.

For everyday tasks, that has largely come true. OpenAI says 1.2 billion people use ChatGPT every week (TNW).

So here is the uncomfortable question about AI affordability. What happens if the cheap end keeps getting cheaper while the powerful end keeps getting dearer? Everyone gets AI. Only some get the best of it, at the volume that changes outcomes.

I’m calling that gap the Intelligence Divide. This piece looks at whether it’s real, how wide it could get, and what might close it.

The economics of intelligence

Every time you ask an AI model something, it runs on chips in a data centre. Producing the answer is called inference. It’s billed in tokens, roughly chunks of words, and each token costs electricity, hardware time and cooling. That is the real cost of artificial intelligence, and it’s paid on every request.

The headline trend is a price collapse. Epoch AI tracked the cheapest way to hit fixed benchmark scores. Performance matching the original GPT-3.5 on the MMLU knowledge test cost $20 per million tokens in November 2022. By October 2024, Google’s Gemini 1.5 Flash 8B matched it for $0.07. Across tasks, Epoch found the price of a given performance level falling between 9 and 900 times a year, and warned that the fastest drops were recent and may not last (Epoch AI, 12 Mar 2025).

That’s the optimist’s case, and it’s strong. Yesterday’s frontier becomes today’s bargain.

The frontier gets dearer faster than prices fall

The catch: nobody doing hard work wants yesterday’s frontier. They want today’s, and today’s costs more to run.

Researchers at MIT FutureTech found that, at a fixed level of performance, frontier models got 5 to 10 times cheaper a year between April 2024 and November 2025. Over the same period, the cost of running the best available model rose by between 3 and 18 times a year (Gundlach et al., arXiv, 2026).

Both can be true. Reasoning models work through long chains of hidden steps before they answer. More steps, better answers, far more tokens. The paper’s own illustration: if token prices fall tenfold but the next 1% of performance needs 100 times more computation, the bill for the frontier rises tenfold. One data point it cites: running OpenAI’s o3 at high effort on the ARC-AGI puzzle benchmark cost about $3,000 per task.

The authors draw the access conclusion themselves. If frontier inference keeps getting dearer, it may concentrate among a smaller set of high-budget users and organisations.

The physical bill

Underneath sits concrete, copper and power. The IEA’s 2026 report on energy and AI says five large tech companies spent more than $400 billion of capital expenditure in 2025, mostly on data centres, with a further 75% rise expected in 2026. Data centre electricity demand grew 17% in 2025, against 3% for global demand. Energy per AI task is falling fast, but more users and agent workloads mean data centre consumption is set to double by 2030, with AI-focused sites tripling (IEA).

That money has to be earned back. Some will come from enterprise contracts and advertising. Some will come from whoever is willing to pay most for the best.

The emergence of the Intelligence Divide

Put those trends together and the shape is clear. A floor that keeps dropping. A ceiling that keeps rising.

The price ladder already looks like that. Free tiers at the bottom, standard plans in the middle, then OpenAI’s $200 and $500 personal plans at the top. Above those sits the business market. In March 2025, The Information reported that OpenAI had discussed agent products at $2,000, $10,000 and $20,000 a month, the top one aimed at “PhD-level research”. Neither the prices nor a launch date were confirmed (TechCrunch, 5 Mar 2025).

So the AI inequality worth worrying about sits above basic access. Almost anyone can use AI. Money buys two things on top: the strongest model, and enough of it to run as agents that work for hours without you.

Is that new? Partly. The wealthy have always had better advisers. What’s different is how fast the premium compounds. A stronger model writes better code, which ships products sooner, which earns the money to buy more compute.

When businesses can afford intelligence that individuals cannot

A household pays for AI out of income. A company pays for it out of expected return. That one difference decides who sits at the top of the price ladder.

If an agent costs a firm $10,000 a month and does work the firm values at more than that, the purchase is easy. For a freelancer, a student or a corner shop, the same price is out of reach. Same model, different arithmetic.

There’s early evidence that AI and wealth inequality are already linked. Using Anthropic Economic Index data, IMF economists Fan and Nguyen estimate AI is saving time worth about $2.7 trillion a year across 86 countries, roughly 3.4% of their GDP. The gains equal about 4.2% of GDP in high-income countries, roughly 0.6% in middle-income ones and 0.1% in low-income ones. Within countries, they tilt towards the highest-paid jobs (VoxEU, 2026).

The same column holds the best counter-argument. Experiments find the biggest productivity gains among less experienced and lower-skilled workers. And the tilt is easing: between November 2025 and February 2026, 42% of countries in the sample saw AI’s gains spread more evenly across occupations. AI can level skills inside a firm while concentrating advantage between firms.

A game nobody can sit out

Treat it as a competitive game. If your rival buys ten times your compute and it pays off, you match them or lose ground. If it doesn’t pay off, they’ve wasted money and you haven’t. For well-funded players, buying in is the safe move. Smaller players can’t match the stake, so their best hope is that the premium tier stops paying off before it buries them.

That dynamic already reaches research. The MIT paper notes that evaluating frontier models on some benchmarks now costs thousands of dollars, which prices out academic groups and hands the job to well-capitalised private firms. When the people checking the claims can’t afford the product, everyone loses a referee.

The impact on education, employment and ordinary people

Students

Students are the clearest test. HEPI’s 2026 survey of 1,054 UK undergraduates found 95% use AI in some form and 94% use generative AI for assessed work. Only 38% said their institution provides AI tools. One of HEPI’s recommendations is that any AI tool that is necessary or advantageous for coursework or assessment should be accessible to all students (HEPI, 12 Mar 2026).

Read that carefully. If the tool matters and the university doesn’t supply it, the student who can pay gets the better version. That’s a grade gap with a direct debit attached.

Workers and freelancers

Two junior analysts with the same title can end up doing different jobs. One uses a free tier. The other’s employer pays for agentic tools. The second is likely to produce more and reach the higher-value parts of the work sooner.

Freelancers and small businesses sit in the awkward middle. They compete with agencies that can run fleets of agents overnight. Their traditional edge was cost, and that edge shrinks when a bigger rival’s extra labour is a model.

Everyone else

For many people, the honest answer is that the free tier is plenty. Drafting a complaint letter, planning a trip or making sense of a council form doesn’t need the frontier. The divide bites where work is complex, long-running or competitive. Unfortunately, that’s where most of the money is.

Will AI become another luxury subscription?

Parts of it already are. A $500-a-month plan is a luxury good by any normal definition.

The more telling signal is the $200 tier. From 30 October, its included Work and Codex usage fell from 20 times the Plus allowance to 10 times, at the same price (TNW). That doubles the effective price per unit of that usage overnight. Sticker prices can hold steady while what you get shrinks. Shrinkflation, applied to intelligence.

Other forces push the opposite way. Labs fight hard for mass-market users, which keeps entry plans cheap and free tiers generous. And mid-range capability keeps falling in price, as the Epoch data shows. A luxury tier at the top doesn’t stop the middle getting better.

So the likely outcome is a tiered market, with every tier open to anyone who can pay its price. The questions that matter are how far apart the tiers drift, and whether the capability at the top is the kind that decides careers and markets.

Can open-source AI prevent this from happening?

Open-weight models, whose weights anyone can download and run, are the strongest counterweight. They’re also closer to the frontier than most people assume.

Epoch AI found that since January 2026, the best open-weight models have trailed the best closed models by an average of four months on its capabilities index (Epoch AI, 29 May 2026). Separately, it found that a single high-end gaming GPU, NVIDIA’s RTX 5090 at under $2,500, can run models matching frontier performance from 6 to 12 months earlier. Epoch adds a caveat: small open models are more likely to be tuned to benchmarks, so the real-world lag may be longer (Epoch AI, 15 Aug 2025).

The MIT paper adds detail. On price against performance, open models dominate the low- and mid-cost range. Closed models still hold the very top.

That’s a real check on pricing power. Nobody can charge $500 a month for last year’s capability when it’s downloadable.

Where open source runs out

Open models shorten the divide in time. They don’t remove it. A four-month lag in a field moving this fast is still a lag, and for competitive work four months is a long time.

Running models locally also needs a top-end GPU and the skills to set it up, which most people don’t have. And the open frontier depends on a handful of organisations choosing to keep publishing weights. That choice can change with one board meeting.

Could governments treat access to AI as an essential digital resource?

The UK has done something like this before. In March 2018, the government laid legislation for a broadband Universal Service Obligation. From 2020 it gave everyone a legal right to request a connection of at least 10Mbps, up to a cost threshold of £3,400 per premises, modelled on the old universal right to a landline. The digital minister at the time called internet access “a necessity not a luxury”, and the government said it expected the minimum speed to rise over time (GOV.UK, 28 Mar 2018).

The pattern is familiar. The telephone and then broadband both started as things the well-off had first. Both ended up with a legal floor once they became necessary to take part in normal life.

Access to artificial intelligence could follow the same route. A floor might mean capable AI provided free to every student through schools and universities, which HEPI’s recommendation already points towards. It might mean stronger models available through libraries and public services, or shared compute for researchers who can’t afford frontier evaluations.

There are good reasons to be wary. Governments buy slowly, and any floor defined today will be dated within a year. State-funded access can lock in a single vendor. And the broadband precedent shows the trap: a minimum that looks generous when it’s written becomes basic very quickly.

Still, the principle travels. If intelligence becomes infrastructure, someone has to decide where the minimum sits.

Who really gets to benefit from the AI revolution?

Two things are true at once. AI is the most widely available powerful technology ever released: more than a billion people use one chatbot every week, and capability that cost $20 per million tokens in 2022 cost seven cents two years later. And the very best AI is getting more expensive to run, more expensive to buy and more concentrated among those who can pay.

The optimists are right about the floor. The pessimists are right about the ceiling. The future of artificial intelligence will be decided in the gap between them.

My expectation is that basic AI ends up close to free, much as search did. The divide opens higher up: in agents that work for hours, in compute budgets that decide who ships first, and in which tools schools and employers choose to provide.

We already accept that money buys better legal advice, better tutoring and faster healthcare. If intelligence becomes one of the most valuable resources in the world, the default is that it joins that list.

The price ladder is still being built. Now is the cheapest moment to decide where the floor should sit, and who pays for it.

Sources

Opened on 31 October 2026.

Implement with AI

Describe what you want the AI to implement or change for this post. It opens with the article and your request already filled in.

Implement with: ChatGPT Perplexity Claude Grok

Keep reading