Cheap routing beats clever prompting
Three places work can run
Any given AI interaction runs in one of three places. On a remote server, accessed through a browser or an application — the default for most chat surfaces. On a local machine that talks to a remote model — the pattern most agentic tools use, with the file system on the user's computer but the reasoning happening elsewhere. Or entirely on the local machine, with a smaller model and no network in the loop. Each has consequences for privacy, latency, repeatability, and cost. None is universally correct; the productivity question is which to use for which category of task.
Tokens as the unit of cost and attention
Almost every commercial assistant charges by the token, where a token is roughly three quarters of a word. Both the input and the output are counted, and output tokens are typically several times more expensive than input ones. A long conversation with attached files can quietly accumulate cost long before any large piece of work has been done. The mature stance is to treat tokens as a real budget. The lever that matters most is not avoiding work but routing it — using the smallest model that will do the job, clearing context that is no longer needed, and being honest about which sessions deserve a more expensive model.
Subscription versus pay-as-you-go
Two pricing patterns coexist. A fixed monthly subscription provides generous usage caps and predictable cost; a metered plan charges per token with no cap but no commitment. The rule of thumb is that anyone using the assistant for several hours a day will reach the subscription's break-even quickly and benefit from no longer thinking about per-message cost mid-flow. Anyone using it occasionally is better off metered. The hidden cost of the metered plan is not the money but the mental tax of watching the meter; that tax distorts which work people are willing to do.
The routing rule of thumb
The single largest source of silent waste in a productivity stack is using a heavy model for work that a light one would handle. A reasonable default is to use the smallest model for classification, extraction, routing, and short replies; the middle model for most writing, analysis, and coding; and the largest model only when a piece of work genuinely benefits from extended reasoning — fewer than one in five sessions for most workers. A routing layer that simply enforces this rule of thumb can reduce blended cost by something on the order of forty to sixty per cent without any visible loss of quality. The savings exceed almost any other tool decision a worker can make.
Privacy, repeatability, and the local option
There is a class of work where the cloud answer is wrong on principle. Anything bound by a confidentiality clause, anything in a regulated jurisdiction, anything where the same input must reliably produce the same output for compliance reasons — these belong on a local model where the inputs do not leave the machine. Local models in mid-2026 are roughly two model generations behind the frontier, which is enough for most extraction, summarisation, and drafting work. They are not enough for frontier reasoning. The mature stack uses both, routed by sensitivity rather than by preference.
If asked to estimate roughly what your AI usage cost over the last month, could you? If not, which is the more likely failure — paying too much, or paying too little for the work it could be doing?