AI Token Pricing 2025: Cost Trends & Sustainability Impact

A million input tokens on a frontier model cost roughly thirty dollars in early 2023. The same class of capability now costs cents, and the cheapest usable tiers sit far below a dollar per million tokens. The collapse is real, but it has not made AI cheap: most bills keep rising, because prompts, context windows and agent loops grew faster than unit prices fell.

This page sets out the list prices per million input and output tokens, how far they have fallen since 2023, five controls that actually cut a bill, and what cheaper tokens cost in energy and in sustainability reporting terms.

How AI Token Pricing Works: Input and Output Tokens per Million

Language model APIs bill by the token, not by the request. A token is a fragment of text, in English roughly four characters. Every provider quotes the same unit: price per million tokens, split into an input rate for everything you send and an output rate for everything the model writes back.

Output almost always costs more than input, usually three to five times more. The reason is mechanical: input is processed in one pass, while output is generated one token at a time, each step re-reading the whole context. Two further rates matter once volume grows. Cached input, where a context the provider has already seen is re-sent, is billed at a fraction of the normal input rate.

Sending a forty-page sustainability report to a model costs roughly 30,000 input tokens. Asking for a two-page structured summary adds perhaps 1,500 output tokens. At an input rate of three dollars and an output rate of fifteen dollars per million, that single analysis costs about eleven cents.

AI Token Price Trajectory 2023 to 2026

Price trackers publish today's rates. What they rarely show is the slope, which is the part that matters when you are signing a two-year budget. The table below follows published list prices per million tokens from the launch of GPT-4 onwards.

Provider list prices in US dollars, standard tier, no batch or cache discount. Prices checked on 4 September 2026.

Read down the input column and the fall is about two orders of magnitude in under two years.

My position on the slope is that it is the least useful number in the table. In my own document-analysis work the per-token price fell far faster than the cost of finishing a job, because the jobs grew: longer context, retries, reasoning steps that bill as output. Budget against cost per completed task, not against the per-token rate, and re-check list prices quarterly.

Cutting AI Token Costs in Practice

Five controls do most of the work. They are listed in the order I would apply them, cheapest change first.

• Cut prompt length. Long system prompts and pasted boilerplate are billed on every single call.
• Use context caching. If the same document or instruction set is sent repeatedly, the cached input rate applies.
• Batch what is not urgent. Overnight classification and extraction rarely need a synchronous answer, and the batch tier is materially cheaper.
• Route by task, not by habit. Extraction and classification run well on the smallest tier. Reserve the expensive model for drafting and judgement.
• Cap output. Set a maximum token limit and ask for structured output.

One buyer's note from consulting practice: the cheapest model per million tokens is often not the cheapest per completed task. A weaker model that needs two retries, a longer prompt and a verification pass can cost more than the capable model that answers once. Before switching a workload to a cheaper tier, measure the full cost of a finished output, including the calls that failed.

What Cheaper Tokens Cost in Energy, Water and Reporting Terms

Falling unit prices have not reduced consumption, they have increased it. This is a textbook rebound: when a token becomes a hundred times cheaper, organisations do not spend a hundredth as much, they run a thousand times more tokens. Aggregate demand, energy use and spend all rise while the price per unit collapses, which is why a finance team can watch the rate drop and the invoice grow in the same quarter.

The physical side is measurable in orders of magnitude rather than decimals. Agency estimates put a single text query to a large model in the range of a fraction of a watt-hour to a few watt-hours, against roughly a tenth of that for a conventional web search. Data centres could account for a materially larger share of global electricity demand by 2030 than the one to two percent they represent today, and water use for cooling scales with the same curve, drawn where the data centres sit rather than where the queries originate.

For a company that buys AI as a service, this is not an abstraction. Purchased AI services sit in the upstream value chain, which places them in Scope 3 rather than in your own energy figure. Organisations preparing a sustainability report under the VSME standard or tracking Scope 3 emissions should ask providers for energy and location data at contract stage, when there is still leverage, and record AI spend as a separate line in ESG data management.

The direction of travel is clear enough to plan around without forecasting a number: list prices keep falling, total consumption keeps rising. Put the price re-check and the energy question on the same quarterly procurement cycle, and the collapse becomes a budget advantage instead of an unmanaged climate risk exposure.

Frequently Asked Questions

Last reviewed: 4 September 2026. List prices change frequently, check the provider pricing page before budgeting.