GlossaryFloor 4 · The Organisationa mosaic: AI is only one tile among manyFloor 4 · The Organisation
cost per token
No. 074 · v2026-08FR: coût par tokenThe cost per token is the unit price of what the model reads and writes: it turns a use into a bill, and it is paid on every call, indefinitely. Like a water meter rather than a subscription: what costs is not the installation, it is the turning of the tap.
What it is not
The cost per token is not the cost of a use. A user who asks a question often triggers several calls, each one re-reading a permanent instruction, documents and a history: the real bill is counted per task completed, not per question asked. Nor is it a stable price: it falls continuously at equal performance, but the uses that appear consume more, so that an organisation’s total spending rises while the unit price collapses.
In depth
Four asymmetries
Four asymmetries explain most of the surprises. Output is billed markedly more than input, which makes long answers and detailed reasoning expensive. Input that has already been seen can be cached and paid at a fraction of the price, provided the prompt has been composed for it. Images consume a large number of tokens. And reasoning models bill their intermediate steps, even invisible ones, which multiplies the cost by a factor that is hard to anticipate since it varies from one run to the next.
The right unit
The right unit for steering is therefore not the price per million tokens but the cost per successful task. That figure incorporates what really counts: the number of calls, the length of the context carried, the retries after failure, and the failure rate itself, since a failed task has cost money without producing anything. Two systems displaying the same model and the same tariff can differ by an order of magnitude on that measure, and it is that measure which decides whether a service is viable.
The levers
The effective levers are few and rank in order of yield. Not calling the model when a simple rule will do remains the first. Using a small model to triage and escalating to a large one only on the difficult cases comes next. Then reducing what is carried, by avoiding sending an entire history on every turn. Finally, structuring the prompt for the cache. Conversely, changing supplier for a few per cent of tariff is almost always the least profitable lever, and it costs team time.
Relations where the neighbours live
Check 3 questions · click your answer
Level 1 · Recognise
Which costs more, at equal volume?
Level 2 · Distinguish
Which quantity has to be steered in order to judge whether a service is viable?
Level 2 · Distinguish
Unit prices are collapsing. Why does an organisation’s bill go up?
Lexigraph, "Cost per token", v2026-08, https://www.lexigraph.org/en/cost-per-token/, CC BY 4.0.