A token is the unit artificial intelligence vendors bill for. Your text is cut into small fragments, roughly 100 tokens per 75 words, and every fragment counts both on the way in and on the way out. In October 2026, a million tokens costs between ten cents and nine euros depending on the model tier. Unit prices have collapsed over three years. Company bills keep climbing. Tandem supports more than 50 companies on AI deployment, and here is what actually makes a budget drift.
What exactly is a token?
A token is a fragment of text, often a syllable or part of a word. Language models do not read sentences, they read this stream of fragments. Your message is converted into tokens on the way in, and the model's answer produces more on the way out. Both are billed, at different rates, with output usually costing four to five times input.
The second concept to keep in mind is the context window. It covers everything the model sees at the moment it answers. Today's message, the conversation history, uploaded files, the instructions of any active assistants, the description of connected tools. All of it is re-read at every exchange. A well-filled window costs money and degrades answer quality.
Why token prices fall while the bill goes up
The cost of a given performance level falls fast. The LLMflation analysis published by Andreessen Horowitz in November 2024 already measured a tenfold drop each year at constant quality. The trend has not stopped since. A performance level worth thirty dollars per million tokens in early 2023 now sits at a few cents. Volume, meanwhile, explodes. Google processed close to 10 trillion tokens a month in 2024. A year later the counter read 480 trillion. By May 2026 it passed 3.2 quadrillion. A factor of 330 in two years.
Finance teams are seeing it. In the index published by Esker in September 2026, 72% of the finance leaders surveyed spent more than planned on their AI projects over the past year. The survey covered more than 300 finance executives across five regions. In the same panel, 65% of CFOs struggle to tie AI use to a specific business result.
The full analysis of this price decline is available from Andreessen Horowitz, which has tracked the index since 2024.
The snowball effect, or why message thirty costs thirty times message one
Here is the mechanism most users never see. At every new message, the model re-reads the whole conversation from the start. Your first message weighs 500 tokens. The thirtieth message in the same conversation weighs more than 15,000, for an equivalent request. The running total reaches 232,000 tokens over thirty exchanges.
The practical consequence fits in one sentence. One topic, one conversation. I worked through the full arithmetic of this compounding effect in my edition on prompting. What else weighs silently, beyond history, are the connectors and apps left switched on. Their description occupies context permanently, even when nobody uses them.
How much does a million tokens cost in October 2026?
The public price lists of the major vendors converge on three tiers. A light tier for everyday work, a mid tier for most business tasks, a frontier tier for complex problems. The amounts below are rounded and converted into euros, taken from vendor pricing pages in October 2026.
| Tier | Input | Output | Cache read |
|---|---|---|---|
| Light | around €0.10 | around €0.45 | around €0.01 |
| Mid | around €2 | around €9 | around €0.10 |
| Frontier | around €9 | around €45 | around €0.90 |
Remember the gap. A factor of one hundred separates the light tier from the frontier tier, on input as on output. An agent running by default on the most powerful model therefore costs a hundred times its frugal equivalent, for a task the light model often handles very well.
Individual subscriptions follow a different logic, with quotas rather than a meter. The detail of Claude plans and their limits sits in our article on Claude subscriptions.
Three reasons an AI budget drifts
The first reason is the arrival of agents. A task handed to an agent triggers a dozen model calls, sometimes more. Where an assistant used one round trip, the agent chains reading, searching, executing and checking.
The second reason is the opening of new use cases. Every price drop makes viable what was not. Cheap electricity filled homes with appliances, and the reasoning holds identically here. Unit prices fall, scope widens, and the bill follows scope.
The third reason is the costliest, and the easiest to fix. Teams keep the most powerful model by default, the one that impressed in the demo. Frontier models are a showcase and unlock the hard cases. In production, a large share of the work runs very well on a fast model, dozens of times cheaper.
The agent reflex deserves challenging before the model reflex. Many needs are met with a configured assistant or a simple automation, far cheaper to run and to maintain. I filmed this pyramid of automation layers, and our business use cases show that the right layer is rarely the most sophisticated one.
Routing and evaluation, the right model in the right place
This is the question our clients ask most often right now. How do you stop token costs from drifting. Our answer fits in two words, routing and evaluation. You test each task across several models, keep the cheapest one that holds the expected quality, and reserve the most powerful for cases that justify it.
Caching deserves a separate mention. Vendors bill already-seen portions of context at a discount, provided the beginning of the message does not move. System instructions, product documentation, a price catalogue, anything stable is re-read ten times cheaper. Change one line in the middle and the cache breaks, so you pay for the whole thing again.
Frugal AI, what the AFNOR reference adds to the debate
Frugality goes beyond budget. The AFNOR Spec 2314 reference, published in June 2024 at the initiative of the French ministry for ecological transition, offers a definition companies can use. An AI service is frugal when the need to use AI, rather than a less energy-intensive solution for the same goal, has been demonstrated.
That definition does executives a concrete service. It moves the question from how to why. Before picking a model and optimising its context, you check that calling a model was the right answer at all. Many of the jobs teams hand to AI are settled by a business rule, a database query or a better-designed form.
The document is free, runs to around a hundred pages and compiles 31 good-practice sheets. It also serves as a methodological base for setting environmental criteria in procurement, which becomes a topic as soon as a purchasing department takes on AI.
The habits that cut consumption by three
Steering does not only happen in the architecture. It happens in team habits too, and that is the cheapest part to fix. Here are the habits that pay off most.
- One topic, one conversation. Every message re-reads the whole history, so a sprawling thread bills the same memory dozens of times.
- Edit rather than correct. Going back to the original message erases the failed attempt. Adding a "no, more like this" keeps it in context and pays for it again.
- Switch off unused connectors. An active connector publishes the description of all its tools at every message, even when nobody calls them.
- Pick the model from the start. Switching model mid-conversation forces a full re-read. Start on the light tier and move up only if quality demands it.
- Upload your files once. A shared project space avoids sending the same document into five different conversations.
- State the length you want. Without instruction, recent models size their answer to perceived complexity, and often run long.
Who is blocked for lack of credits?
This is the most interesting question, and almost nobody asks it. When a company moves to usage-based billing, the first instinct of management is to cap. Cutting everywhere gives the worst possible outcome.
The people who run out of credits are often the most advanced. They are the ones who industrialised a flow, built a shared assistant, or replaced three hours of manual work with twenty minutes of supervision. Throttling them punishes exactly the behaviour you were trying to install.
The right answer is a split. A baseline for everyone, enough to explore without thinking about it. A wider envelope for AI champions and for agents that produce a measurable result. That split is decided in committee, with an owner per use case, as described in our article on AI governance. Enterprise platforms such as Dust expose consumption per user and per assistant, which makes that arbitration possible without a homemade spreadsheet.
Should you worry about the cost of AI tokens?
No, provided you move from a subscription mindset to a steering mindset. A flat fee is endured. Usage-based consumption is steered, and it is steered with three numbers only. Cost per automated task, the share of the budget going to the frontier tier, and the number of people blocked by a cap.
The underlying trajectory stays favourable to companies. The price of a given quality level will keep falling, and the tasks you pay dearly for today will cost a fraction of that in eighteen months. The danger was never the price of a token. It is the lack of visibility on what your teams do with them.
The starting point is an inventory of use cases and their real cost. That is exactly what we frame in an AI audit, within a few weeks, with the routing grid that goes with it.
