What is a token in AI pricing?

A token is the unit AI models are metered in — roughly three quarters of a word. "Tokenomics" is two tokens ("token" + "omics"). You are billed for tokens going in (your prompt) and tokens coming out (the answer).

Like electricity: you pay for what you pull. API access is metered per million tokens, while consumer products bundle an allowance into a flat monthly subscription. The two are different units and should never be averaged together.

Is the cost of AI going up or down?

Per-token prices for a fixed level of capability have fallen sharply and continue to fall, while frontier capability keeps its price. Total spend still rises because usage grows faster than unit cost falls.

What does one AI image or video actually cost?

At published API rates, a text answer costs a fraction of a cent, a generated image costs cents, and eight seconds of generated video costs on the order of tens of cents to a few dollars depending on the model. Tomorrow tracks these per-task costs on the Tokenomics page.

AI Tokenomics

Name: AI Tokenomics — cost per token, per task and per outcome
Creator: Tomorrow

The unit economics of intelligence

AI is sold the way electricity is sold: you pay for what you pull. This page tracks that meter — what a unit costs, where it is falling, and what a finished job really costs.

Metered, like electricity

Every job has a price you can read off a meter.

The unit is a token

Roughly three quarters of a word, counted in and out.

A plan is not a quantity

A $20 subscription is a bundle, never a rate.

Three units, never mixed

Most confusing AI pricing arguments are two people quoting different units at each other. Everything on this page is labelled with one of these three, and they are never averaged together.

USD / 1M tokens

What a unit of output costs

Published API list prices, input and output blended. This is the meter reading — the electricity rate of intelligence.

Frontier work runs dollars per million tokens; commodity tiers run cents.

Tokens / $1

How much a dollar buys

The same rate read backwards, which is how routing decisions actually get made: send the easy 80% of tasks to a lighter model.

The spread between tiers is the size of the lever, often 50x or more.

never averaged

USD / seat / mo

What a bundled product costs

The cheapest paid plan of a product, per person. A seat buys a product, not a quantity — prompt length never shows up here.

Typical tracked plans cluster near $18–$25 a month.

01 and 02 — one family: the same meter, read two ways

03 — a bundle, not a quantity

What does it actually cost to do a thing?

The meter reading, in plain terms. Every job below is priced off published list rates and set next to something you already pay for: an hour of television costs about 1.7 cents. Switch the model to see how much of the bill is a choice rather than a fact.

Metering on GPT-5.6 Terra · 0.25¢ per 1k words in

0.65¢

Ask a question

A short question, a couple of paragraphs back

≈ 23 minutes of TVmetered

1.4¢

Write a recipe

Ingredients, method, substitutions — about a page

≈ 50 minutes of TVmetered

1.6¢

Translate a page

One page of text into another language

≈ 55 minutes of TVmetered

8.8¢

Summarise a 40-page PDF

Read the whole document, give back one page

≈ 5.2 hours of TVmetered

$1.07

Build a small feature

An agent reading files, writing code, re-running itself

≈ 63.2 hours of TVmetered

4.0¢

Generate an image

One high-quality 1024px image

≈ 2.4 hours of TVflat rate

9.0¢

Read 1,000 words aloud

Text-to-speech, about eight minutes of audio

≈ 5.3 hours of TVflat rate

$3.20

Generate 8s of video

One short generated clip with audio

≈ 188.2 hours of TVflat rate

Costs are the published list price of the work itself — what the vendor charges to run the job, before anyone adds a subscription, a margin or a monthly plan on top. Television equivalents use the US average residential rate of $0.17 per kilowatt-hour and a 100-watt set. Flat-rate items are billed per image, per character or per second, so the model selector above does not move them.

The scissors: floor collapsing, ceiling rising

One chart, two opposite truths. The commodity floor is the cheapest price at which GPT-4-class capability can be bought — it has fallen by orders of magnitude. The frontier ceiling is the price of the best model on the market — it has gone up. Averaging them into a single 'AI price' hides the entire story.

Commodity floor

$0.138

GPT-5 nano · cheapest GPT-4-class blended rate

Frontier ceiling

$16

Claude Opus 5 · best available model

Spread

116×

Frontier costs this much more than commodity capability

Commodity floor (GPT-4 class) Frontier ceiling (best model)

The floor has fallen 95% in the last 12 months while the ceiling has held. Hover a point for the model that set it.

Take the meter with you

Tokenomics is one instrument in Tomorrow, the emerging tech index. Watch it, pull it, or read how it is verified.

WatchTokenomics, the showThe same index read out loud every week — what moved, what it costs, and why.Use the dataAPI & terms of useEndpoints, field definitions, fair use and licence. No key, no sign-up.Go deeperHow the numbers are madeSources, daily refresh and the three-lab check every claim passes before it ships.

Looking for where the market is moving rather than what it charges? That lives in Market Pulse.