Intelligence is getting cheaper. Our expectations aren’t.

In March 2023, the best AI model you could buy was GPT-4, at $37.50 per million tokens. Eighteen months later, a model just as good cost thirteen cents. That is 286 times cheaper.¹

If cars did that, a ₹10 lakh car would cost about ₹3,500 a year and a half later.

And yet hardly anyone feels that AI got cheap. Simon Willison, a programmer who writes about AI models almost every day, pays $200 a month for two AI subscriptions. At list prices, his last thirty days would have cost $2,180.² Companies feel it too. This year Uber spent its whole annual AI budget in four months, and Salesforce expects to pay Anthropic about $300 million.³

So which is it? Both. Intelligence keeps getting cheaper. What we expect of it keeps getting more expensive. The moment a level of intelligence becomes cheap, we stop wanting it. We want the next level, and a lot more of it.

What you are paying for

AI companies charge by the token, a chunk of text that is usually part of a word. “Cheap” is one token; “unaffordable” is three. In English, a thousand tokens is about 750 words.

Prices are quoted per million tokens, with one price for what you send and a higher one for what the model writes back. When this note gives a single price, it blends three parts input to one part output, a common convention among price trackers.⁴

Part one

Intelligence is getting cheaper

Here is that drop step by step. At each point since GPT-4 launched, this is the cheapest model, open or closed, that does at least as well.

In everyday terms: having GPT-4 proofread a 300-page book in 2023 cost about $11. With GPT-6 Luna, which scores far higher, it costs about seven cents.⁵ Then something more interesting happened: hardly anyone took the savings. When a better model comes out, most teams move to it within weeks instead of saving money on the old level.⁶

Switch the chart to “Every level” and you see the pattern repeat. In Epoch’s tests, the first model to score 65% on GPQA Diamond, about what PhD experts score in their own fields, was OpenAI’s o1 in December 2024, at $26.25 per million tokens. Four months later Alibaba’s open model QwQ-32B did it for about 16 cents on DeepInfra, a hosting service. The first model to pass 90% cost $4.50 in November 2025; eight months later DeepSeek’s V4-Flash did it for about 18 cents. GPT-4’s level kept falling too: by August 2026 Alibaba’s Qwen3.7-Flash beat it for about five and a half cents, nearly 700 times less than GPT-4.¹

Epoch AI, a research group that tracks this closely, measured the full cost of reaching a given score, including all the thinking a model does before it answers. Across five benchmarks it falls about 47% every quarter, roughly thirteen times a year. By their reckoning that is six times faster than the long-run fall in the cost of computing, and eighteen times faster than batteries.⁷

Smarter and cheaper at the same time

Holding quality fixed undersells it, because quality didn’t stay fixed. Plot 148 model versions from 13 labs, open and closed, by how capable they are and what they cost, and each year’s best options sit higher and further to the left than the year before. Whatever you spent, you could buy a better model than a year earlier.

Here is the same thing as a report card: GPT-4 next to the smallest and the largest of OpenAI’s newest models.

Why it keeps falling

Big models teach small ones. A small model is trained to imitate a large one, and ends up nearly as good for a fraction of the cost. Google says it made Gemini 1.5 Flash this way.

Models use only the part they need. DeepSeek-V3 has 671 billion parameters but uses about 37 billion for each token, like a hospital that sends you to a few specialists instead of every doctor.⁸

Repeated work gets reused. Instructions that arrive again and again are processed once and stored, and OpenAI charges as little as a tenth of the normal price for them.⁹

Competition. In January 2025 DeepSeek’s open model R1 came close to OpenAI’s o1 at about a twenty-seventh of the price. Nvidia lost $589 billion of value in a day, and that June OpenAI cut the price of o3 by 80%.¹⁰ Open models add to the pressure: anyone can run them, so hosts compete on price. A day after OpenAI released its own open model, gpt-oss-20b, DeepInfra was selling it for about 7 cents.¹¹

Part two

Our expectations aren’t

If prices fell this fast, why are bills going up? Because nobody buys a fixed level of intelligence. We buy the newest good model, we use it as much as we now want to, and the best model is never the cheap one.

In 1865 the economist William Stanley Jevons noticed the same thing about coal. Engines got far more efficient, and Britain burned more coal than ever, because cheap power found new uses. “It is the very economy of its use which leads to its extensive consumption,” he wrote.¹²

Light is the clearest example. In Britain it became more than 3,000 times cheaper between 1800 and 2000, and each person used 6,500 times more of it. So people ended up spending more on light, not less.¹³

AI is doing the same thing in about three years, in four ways.

We want the newest model

GPT-4-level intelligence now costs almost nothing, and hardly anyone settles for it. ChatGPT dropped GPT-4 itself in April 2025.⁶ The labs know this, and they price the top of the range for it.

It isn’t only OpenAI. Anthropic’s Claude Fable 5 launched at $10 per million input tokens in June 2026, twice the price of Opus, and Google says Gemini 4 Argon will double in price on 1 December. Even the cheap lines sometimes climb: Claude 3.5 Haiku cost four times what Claude 3 Haiku did, and Gemini 3.5 Flash cost three times its predecessor.¹⁴

We want it to think

Newer models work through a problem before they answer, writing out pages of reasoning you never see, and you pay for every word of it. When Google moved from Gemini 2.0 Flash to 2.5 Flash with thinking switched on, the new model used about seventeen times as many tokens on Artificial Analysis’s tests, and each one cost more, so the same tests cost about 150 times as much.¹⁵ A cheap price per token means little when every answer takes seventeen times as many.

We want it to work for hours

In 2023 we asked AI for a paragraph. Now we hand it work that would take a skilled person most of a day. METR, a research group that tests this, measures the longest task a model finishes half the time. For GPT-4 it was about four minutes. For Claude Opus 4.6, in February 2026, it was about twelve hours.¹⁶

Longer tasks eat tokens. An agent fixing a bug reads files, runs the tests, tries again and reads more, so one request from us becomes hundreds of calls to the model.

We want it everywhere

AI is now inside search results, email, documents and phones. Google says 2.5 billion people a month see its AI summaries in Search, far more than the roughly one billion who use the Gemini app.¹⁷

Put those together and the bill rises while the price collapses. Take one job, answering a million customer questions, and do it the way we did in 2023 and the way we want it done now.

That is how Simon Willison’s $200 of subscriptions came to use $2,180 worth of tokens in a month. Gartner predicts that by 2028 companies will spend more on AI coding per developer than the average developer’s salary.¹⁸ That isn’t AI getting more expensive. It’s us asking more of it.

The fine print

Same price, smaller pack. Claude Opus 4.7 kept its list price, but a new way of cutting text into tokens turned the same English text into about 40% more tokens in Simon Willison’s tests. Anthropic itself says up to 35%.¹⁹ Shrinkflation, applied to words.

Cheap doesn’t always last. In August 2026 DeepSeek raised the price of its Flash model about threefold for input and almost fivefold for output at peak hours.²⁰ Even off-peak, DeepSeek’s cheapest model now costs more than GPT-6 Luna.

Flat plans were a subsidy. “Unlimited” plans are becoming usage bills and weekly limits; GitHub moved its Copilot plans to usage billing on 1 June 2026.²¹

Cheap per answer, not in total. By May 2025 a typical Gemini prompt used about 0.24 watt-hours, 33 times less than a year earlier, yet the IEA expects data centres’ electricity use to roughly double between 2025 and 2030, to about 950 terawatt-hours, more than all of Japan uses today.²²

What this looks like from India

Picture a helpline that answers farmers’ questions about crops, weather and prices, in Hindi, on WhatsApp. Each answer reads a few thousand tokens (the question, the conversation so far, the local advisories) and writes a few hundred. At GPT-4’s 2023 prices, one answer would have cost about ₹11. At today’s prices for a far better model, it costs about five paise. If a million farmers asked ten questions a month, that is the difference between about ₹11 crore a month and about ₹5 lakh.²³ In 2023 that was out of reach for almost anyone. Today a small team could afford it.

Three more things look different from here.

Indian languages pay a tax. The same sentence in Malayalam or Tamil is cut into more tokens than in English, so it costs more. On OpenAI’s older tokenizer, Indian languages needed about eight times the tokens of English on average, and Malayalam thirteen. OpenAI’s newer tokenizer cut the average to about 2.1 times, and Bengali still needs roughly twice as many tokens as English.²⁴

“Cheap” depends on where you stand. A $20 plan is about 0.3% of America’s monthly GDP per person, and about 9% of India’s.²⁵ Falling prices matter more here than anywhere.

We have seen this film. Mobile data in India fell from about ₹269 a gigabyte in 2014 to about ₹10 by 2022. We didn’t save the difference; we went from barely using data to about 31 gigabytes a month each. And once the price war ended, tariffs went up again in 2024.²⁶ OpenAI’s free year of ChatGPT Go for Indian users began on 4 November 2025, so the first renewals at ₹399 a month start in November 2026.²⁷

What to do with this

If you build software, plan for what today’s best model does to cost about a tenth as much within a year, and for your usage to grow faster than that. The skill that matters now is routing: send each step to the cheapest model that is good enough, and save the expensive one for the steps that need it.

If you buy it, decide whether you need the newest level or a known one. A known level of intelligence is close to free. The very top is getting pricier.

For everyone else, the question isn’t whether AI is getting cheaper. It is. The question is what we will ask of it next.

Intelligence is getting cheaper every month. What we want from it is growing faster. Only the second shows up on the bill.

If this was useful, share it on X, or write to me with the numbers from your own bill.