Token economics in 10 minutes

Humans read technical text at roughly 150 words per minute, which comes to about 1,500 words in ten minutes. This post is just under 1,500 words.

1Introduction

This post was inspired by the growing number of headlines about token consumption and spend. It intends to explain what tokens are, how they translate into spend, how to manage them, and what companies are doing about them.

As a disclaimer: none of the content in this post is original. It is a synthesis of content produced by the FinOps Foundation and the Tokenomics Foundation.

2What a token is

A token is the unit in which a model reads and writes. Large language models do not read words and sentences; instead, they translate them (and audio, images and video) into strings of numbers, one number per token.

Different models tokenise in different ways. Below is an example from the tokeniser behind OpenAI’s GPT-5.x and o-series models, which splits this 54-character sentence into 10 tokens, each shaded block being one token. You can try it yourself at platform.openai.com/tokenizer.

Tokens have become the basis for pricing for broadly two reasons:

1. They are agnostic of their modality; it does not matter what the LLM consumed or produced; it can be expressed as tokens.
2. They map well to the economics of compute. Much as a kilowatt-hour of electricity maps back to work done at the power station, tokens map back to work done by GPUs.

Tokens end up produced and sold like a metered commodity: the provider generates them, you buy them, and what you do with them is up to you.

3Two kinds of token

There are two broad categories of tokens: input tokens and output tokens.

Two key things to note:

1. Your output tokens become your input tokens. Every turn of a conversation replays the history back into the model, so what the model wrote last turn is billed again as input this turn. This is why long conversations and long-running agents get more expensive with every turn.
2. “Thinking” counts as output tokens. Reasoning models produce a chain of thought on the way to an answer. Most of it never appears in the response you receive, but you are charged for all of it at the output rate.

4Why output tokens cost more

Model providers price input and output tokens differently, generally with the former being cheaper than the latter.

The asymmetry is not arbitrary; it comes from the mechanics of inference:

• “Reading” is cheaper because the model can read all inputs in parallel (this is referred to as prefill, and is compute-bound).
• “Writing” is comparatively more expensive, as the model needs to generate each output token one at a time (this is the decode stage, and is memory-bandwidth-bound and the main bottleneck of inference on current hardware).

5Translating tokens into dollars

Token spend is the product of token volume and unit cost. Beneath those two factors sit ten levers:

6A quick detour: goodput

One lever worth touching on from the tree above is latency tier: the speed at which tokens come back to you, and the premium or discount attached to that speed. The FinOps Foundation’s writing on token economics³ introduces a key term:

A token in isolation is a pure commodity; what differentiates one token from another is the speed at which it is delivered. That gives inference three broad regions:

• Bulk tokens: cheap to serve but unsuitable for interactive use. This is what long-running agentic loops without a human in the loop will consume.⁴
• The Goldilocks zone: moderate interactivity, the sweet spot for general-purpose inference, and where most workloads sit today.
• Premium low-latency tokens: guaranteed time-to-first-token for time-sensitive use cases, priced accordingly.

It is likely that time-to-first-token, sustained tokens-per-second-per-user, and similar metrics will become part of enterprise agreements (and potentially consumer offerings). For example, Anthropic already offers a fast mode for Opus, which provides the same model but served at up to 2.5× the output speed — priced at $10/$50 per million input/output tokens, against $5/$25 for the standard version.

³FinOps Foundation, “Token Economics: The Atomic Unit of AI Value” — finops.org/insights/token-economics-the-atomic-unit-of-ai-value

⁴Ben Thompson, “The Inference Shift”, Stratechery, May 2026 — stratechery.com/2026/the-inference-shift. On why agentic inference is a distinct market: “lower speed isn't nearly as important a consideration if there isn't a human in the loop.”

7There is more to AI spend than just tokens

Tokens form a large part of the mainstream conversation on AI spend, especially at the level of the individual employee, but AI spend is much broader than token spend:

Now that we understand tokens, let’s discuss some of the real world implications.

8The price environment in 2026

The FinOps X keynote in June 2026 highlighted two pricing trends:

1. Model providers are moving from flat seat-based pricing towards seat pricing plus pay-as-you-go for consumption. This is a structural shift for many organisations that are familiar with traditional seat-based pricing for SaaS and now need to begin forecasting a variable cost. An analogy would be if Microsoft began charging for PowerPoint based on how many slides you created.
2. Per-token list prices are still declining, but the character of the decline has changed: the rate has slowed, and the declines are concentrated in the commodity tiers rather than at the frontier.⁵

⁵I flag this one as the keynote's claim rather than mine. Whether token prices are declining depends on how you measure: the price of a fixed model over time, the list price at a capability tier, or the price per task. These give different answers; by the second measure, the cheapest tiers have in places gotten more expensive.

9AI spend management is an emerging industry

The speed at which AI has become integrated into businesses, combined with the unpredictable nature of its cost structure, has led to the emergence of an entirely new industry:

1. Open standards are being created: the FinOps Foundation’s inaugural writing on token economics³ (the core source for this post) was published alongside the launch of the Tokenomics Foundation, whose remit is to create a shared, open-source standard for managing AI spend.
2. Advice on token economics has already become a professional service, with consultancies extending their cost-out offerings to cover spend management and optimisation.⁶
3. Platforms such as SAP⁷ and Microsoft Foundry⁸ are building the tooling for managing token spend.
4. Companies are investing in the capability in-house, as evidenced by the token spend management frameworks Adobe and Pinterest presented during the FinOps X keynote.

⁶One example: Deloitte, “How to navigate the economics of AI” — deloitte.com/global/en/services/consulting/perspectives/how-to-navigate-economics-of-ai.html

⁷SAP Community, “BTP FinOps: keeping track of your credits” — community.sap.com/t5/cloud-finops-ideas/btp-finops-keeping-track-of-your-credits/ba-p/13737793

⁸Microsoft, “Build 2026: from observability to ROI for AI agents on any framework” — devblogs.microsoft.com/foundry/build-2026-from-observability-to-roi-for-ai-agents-on-any-framework

10Practical application

If you use coding agent tools (e.g. Claude Code, Codex), it is very likely you run them via a subscription and have not needed to contend with managing your token use or spend. A way to make the concepts in this article concrete is to analyse your own local usage logs through a tool such as ccusage⁹. It is an open-source tool, available via GitHub, which reads your token consumption from local logs and translates it into what it would have cost had you been operating on a pay-as-you-go basis.

For example, my usage on the Claude Max plan would have equated to the following had I paid via consumption:

⁹ccusage — github.com/ccusage/ccusage