Token economics is product design

Price per token is not the bill

Model pricing tables encourage a simple comparison: input rate, output rate, context window. The actual cost of agentic work depends on the system sending the request.

Repeated tool output can dominate input. An unstable prefix can prevent provider cache reuse. A frontier model can spend most of its time classifying routine work. Parallel agents can multiply the same context. A long conversation can carry irrelevant history because deleting it feels risky.

Miton calls the response to those problems token economics. The objective is not always the smallest bill. It is the best useful outcome for each cent under the user’s quality, privacy and provider constraints.

Four linked decisions

Choose an eligible model for the task. Routing begins with models the user permitted, then uses task, capability, health, quality and cost signals. A stronger model can be recommended when the work deserves it; routine work need not consume it by default.

Shape context rather than dumping it. Tool output is bounded. Repeated results are removed. Older turns are condensed according to task needs. Relevant project knowledge is assembled instead of replaying every historical token.

Preserve cache-friendly prefixes. Stable instructions and tool definitions remain ordered so providers with prompt caching can recognise work they already processed. Cache behaviour is a dynamic property of the provider and model, so the harness must adapt rather than assume one universal rule.

Show spend where decisions happen. Per-turn cost, model use, project totals and caps let the user connect behaviour to money. A cost dashboard after the invoice is useful analysis; a visible meter before the next dispatch changes the outcome.

Why bring-your-own-model economics matter

When the workspace sells token credits, it benefits from opacity and margin. Miton separates the workspace licence from model spend. Hosted charges remain on the provider account the user chose; local inference has no hosted per-token charge.

That alignment gives the workspace a clearer job: reduce waste, improve routing and make every new model more useful. Model economics will keep changing. The system should turn those changes into better outcomes rather than a new bundle name.

There is also an access argument. The capability is already cheap; the important question is what the technology is for. A workspace should not require a $50 monthly subscription before someone can learn, build, write or automate. Local inference, free provider tiers and bring-your-own hosted models are different trade-offs, but they let the cost of the model remain visible and chosen by the person doing the work.

That does not make inference free or every model suitable. It means the product should not turn model access into a hidden margin, pretend that a bundle is neutral, or force a user to pay for frontier capacity when a cheaper eligible model would do. Cost-aware design is partly accounting and partly a decision about who gets to participate.