AI prices did not fall. They fanned out.

We tracked the published API price of 86 models across 10 providers, back to the day GPT-4 launched in March 2023, with a first-party source on every single number. The data does not show a tidy collapse. It shows a fan opening. The cheap end fell through the floor. The expensive end climbed past anything that existed in 2023. And the gap between them blew open until "the price of AI" became a number that describes nothing.

Here is the figure to keep. Right now, a million tokens of model output costs between $0.10 and $600 depending on which model you call. That is a 6,000x spread, for the same unit, at the same moment. There is no single price of intelligence. There is a price list with four orders of magnitude on it.

Coverage

86 models, 10 providers

Cheapest input fell

857x

Input spread today

4,286x

Output spread today

6,000x

The method, up front

One paragraph, because the whole thing rests on it. Each row in the dataset is one price observation: a provider, a model, input or output tokens, the dollar price per million tokens, the date it took effect, the date we last checked it against the official page, and a link to that page. We measure the standard published list price, not promotional, batch, or cached rates, because list price is the number you get quoted before you optimize. It is all open, CC BY 4.0, and one pull request away from being corrected if we got something wrong.

The floor fell through the floor

In March 2023, the cheapest way to get a million tokens of frontier output was GPT-4, at 60 dollars. There was no budget tier. That was the price of the good model and the price of the only model. Today the cheapest million input tokens in the dataset costs $0.035, and the cheapest output costs $0.10. The cheapest input price has dropped 857x since the GPT-4 launch.

Take a real workload: 10 million tokens in, 2 million out, a meaty day of coding-agent traffic. On GPT-4 at 2023 prices that job cost about 420 dollars. On the cheapest tier today it costs under a dollar. This is the part everyone already wrote about, so we will not dwell. Competition from open-weight models, distillation, cheaper hardware, and a wave of small models drove the bottom of the market toward zero. True, well covered, move on. Because the moment you look at the other end of the fan, the tidy story breaks.

The flagship tier did not get the memo

If prices fell 10x a year across the board, the flagship tier should have fallen with everything else. It mostly did not, and two of the biggest labs went in opposite directions.

Anthropic cut its Opus flagship once, in a single step, after holding it flat for over a year. OpenAI's flagship line did the opposite: it got more expensive over its own lifetime, quietly quadrupling its input price in eight months while the headlines said AI was getting cheaper. The cheap tier is not safe from this either. Anthropic's budget Haiku is about four times more expensive than the cheapest Claude used to be. None of these are hikes you would find by checking today's pricing page. The old number is gone.

A ceiling appeared that did not exist in 2023

In 2023 the most expensive model you could call was GPT-4 at 60 dollars per million output tokens. That was the ceiling. Today the ceiling is OpenAI's o1-pro at $600 per million output tokens, ten times higher than the priciest option that existed three years ago. Premium reasoning tiers, models that think for a long time before answering, opened a whole new band of the price list above where the old frontier used to sit. So while the floor was falling hundreds of times over, the ceiling rose about 10x. The fan opened from both ends at once. That is why the spread is 6,000x and not, say, 50x.

Why this happened

Three forces, pulling apart. At the bottom, intelligence became a commodity and got priced like one: open weights set a floor near the cost of the GPUs, and everyone selling a small model has to price against "or you could just run it yourself." At the top, providers stopped competing on flagship sticker price and started competing on capability per token, so the best model has no pressure to get cheaper, and reasoning models gave providers a reason to charge more because they burn more compute per request. In the middle, the unit of value quietly changed: the price per token is falling, but the best models now spend far more tokens per task, thinking out loud before they answer. Cheaper per token, more tokens per job. Your bill is not tracking the chart that goes down and to the right.

Nobody kept the receipts

Here is the uncomfortable part. We could only assemble this by reconstructing it from first-party pages and archives, after the fact. Provider pricing pages have no changelog. When a price changes, the old number is overwritten and gone. When a model is retired, its pricing page often disappears with it. The single most-cited stat in this whole space, the famous "1,000x cheaper," was built by scraping the Wayback Machine, because the primary record had been deleted.

Think about what that means. The industry that bills you by the token does not keep a public, dated, sourced history of what it charged. If you want to know what a job actually cost the day you ran it last spring, the authoritative answer no longer exists on the vendor's own site. That is the gap this dataset exists to close. Every price has a date and a source. Going forward, a bot diffs each provider's official page daily and a human verifies every change before it lands, so the next quiet repricing gets caught the day it happens instead of being reconstructed from an archive two years later. It is the changelog the pricing pages should have shipped and never did.

What it unlocks

The point of keeping receipts is that you can answer questions the live pricing page cannot. You can reprice a historical bill at the rates that were actually in effect on the day, instead of extrapolating from today's number. You can audit a vendor's "we cut prices" announcement against the date the API page actually changed, and find the cuts that shipped with no announcement at all. You can plot the fan yourself, because the whole thing is a CSV. (Point-in-time repricing is also why we built it: it is the engine behind Goei, our cost dashboard.)

The tidy line was never wrong, exactly. It was just an average drawn across a market that stopped having a single price, and the average of cents and hundreds of dollars is a number that has never appeared on anyone's invoice. The receipts are messier than the line. They are also true, and you can check every one of them.