Why AI prices keep falling, and what that means for you

Two years ago the advice was to use the cheapest model that could do the job because the good one was too expensive to run at scale. The advice still holds. What has changed is how good the cheap one is.

The numbers, as published

Anthropic's list prices, per million tokens, in September 2026:

• Fable 5.1: 10 dollars in, 50 dollars out
• Opus 5.5: 4 dollars in, 20 dollars out
• Sonnet 5: 2 dollars in, 10 dollars out
• Haiku 4.5: 1 dollar in, 5 dollars out

And the sentence that matters more than any of those lines: Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work while costing 40 percent less to run than Opus 5. A tier of capability moved down a tier of price in one release.

Three reasons prices fall

1. Efficiency. The same answer with fewer tokens and less compute costs less, and most of the engineering effort between releases goes here rather than into raw intelligence.
2. Competition. When two or three companies sell something similar, nobody gets to hold a price. Every launch is partly an answer to someone else's launch.
3. Volume. Inference costs fall as the hardware fills up. A model that is busy is a model that is cheap to serve.

None of these is generosity, and reading them as generosity leads to bad decisions. The price fell because it could, and it will rise again in any market where one company stops having a rival.

What it means if you are buying

• Do not sign long contracts at today's price for a capability that will be cheaper in six months.
• Do not architect around one vendor's model name. Keep the model behind a setting you can change.
• Do measure. The cheapest model that passes your own test is the right one, and the test has to be yours.

What it means if you are only using these things

The subscription prices have been much stickier than the API prices. Claude Pro is 17 dollars a month on the annual plan and 20 monthly. ChatGPT Plus is 20 dollars a month. Those numbers have barely moved while the underlying cost of running a model has fallen hard.

What you get instead of a price cut is more for the same money: better models on the same plan, more generous free tiers, and offers aimed at students and teachers. If you want to feel the price war as a consumer, that is where to look for it.

Offers on the tools we track

A prediction, clearly labelled as one

Everything above is sourced. This paragraph is not, so treat it as opinion. The next thing to get cheap is long running work: models that go away for an hour and come back with something finished. That is the expensive kind of use today, and it is where each release is putting its effort, which is usually where the next price cut lands.

What a price cut looks like on a real bill

Take a support team running 20,000 customer questions a month. Each one sends about 2,000 tokens of context and gets back about 400.

• That is 40 million input tokens and 8 million output tokens a month.
• On Opus 5.5 at 4 and 20 dollars: 160 dollars of input and 160 of output, so about 320 dollars.
• On Sonnet 5 at 2 and 10 dollars: about 160 dollars.
• On Haiku 4.5 at 1 and 5 dollars: about 80 dollars.

The gap between the top and bottom of that list is 240 dollars a month, which is real but small. Now multiply the volume by fifty, as a company of any size does, and the same choice is worth 12,000 dollars a month. This is why model selection stops being a preference and becomes a budget line.

Why subscriptions have not fallen with the API

API prices are a commodity market and behave like one. Subscriptions are a product, and products hold their price while adding features, because 20 dollars a month is a number people have already accepted.

So the consumer version of a price war looks different. It shows up as a better model on the same plan, a more capable free tier, and offers aimed at groups worth capturing, which is exactly why the student and teacher offers exist.

What would make prices rise again

1. Consolidation. Two serious competitors keep prices honest. One does not.
2. Compute scarcity. Inference is cheap when hardware is available, and that has not always been true.
3. Regulation costs, which land on the provider and travel to the invoice.

None of those is imminent as of September 2026, and none is unimaginable. Build so you can change model, and the question stops being urgent.