Frontier models are not cheap. Cheap demos are.

The cheapest AI system is the one that only has to work once.

A clean prompt goes in. A polished answer comes out. Somebody checks the usage dashboard and discovers the model call cost less than lunch. The conclusion writes itself: intelligence is now almost free.

What the receipt does not show is how carefully the demonstration was prepared. The source material was selected, the task was friendly, the output was inspected by a person and the awkward attempts disappeared before the meeting. One successful response is being priced as if it were a production system.

Real revenue work does not arrive as one clean request. It arrives as incomplete CRM records, conflicting customer evidence, ambiguous company names, changing offers, permissions, exceptions and deadlines. The system has to do more than generate. It has to find, interpret, decide, act and show why the work can be trusted.

A cheap answer and a cheap system are not the same thing.

The model reads the context again every time the system thinks.

Model pricing is usually presented as a rate per unit of input and output. That makes the arithmetic look wonderfully small. But an agentic system may read a long instruction set, retrieve business context, call a tool, inspect the result, revise its plan and try again. Each turn adds work. Longer context and deeper reasoning can make a single business task many model calls rather than one.

Then the surrounding services join the bill: search, enrichment, retrieval, transcription, image generation, data movement, storage and the software that coordinates them. A failed tool call may need recovery. An uncertain answer may need a second model or a human review. A high-consequence action may need both.

None of this makes frontier intelligence uneconomic. It means the unit being purchased is not a token. The useful unit is a successfully completed task at an acceptable level of quality, speed and risk.

• Context consumed across every model turn.
• Reasoning, retries and alternative paths.
• Retrieval, search and third-party tool calls.
• Evaluation, review and exception handling.
• Monitoring, maintenance and improvement after launch.

The cheapest model can produce the most expensive outcome.

A smaller model that costs less per call may need more retries, more elaborate instructions and more human correction. If it misses a buying signal, invents a customer claim or routes an account badly, the model saving is irrelevant beside the commercial consequence.

The opposite mistake is to use the strongest model for every step. A frontier model should not classify a field that a rule can validate or rewrite text that a smaller model handles reliably. Capability should follow consequence. Use ordinary software for fixed rules, economical models for bounded work and frontier models where difficult interpretation changes the outcome.

This is model routing as a business decision. The question is not, ‘Which model is cheapest?’ It is, ‘What is the least expensive system that completes this class of work dependably?’

Price the path to accepted work.

A useful cost model follows the task from demand to acceptance. It includes what happens when the first attempt is weak, the case is unusual or the system needs to learn. This turns an unpredictable technology bill into an operating question the business can manage.

Budget for the outcome before optimising the token bill.

Choose one recurring revenue task and establish its current cost, quality, cycle time and failure rate. Then run representative cases through the proposed system. Include easy work, messy work and the exceptions that carry commercial risk. Divide the complete operating cost by the number of outputs the team would actually accept and use.

Only then optimise. Cache repeated context. Route simpler steps to smaller models. Replace interpretation with rules where the answer is fixed. Reduce unnecessary turns. Improve the evidence so the system does not spend money reasoning around missing information.

Frontier models can create extraordinary leverage. Pretending they are free makes that leverage harder to operate, price and trust. Serious buyers do not need the lowest demo cost. They need honest economics for useful work at scale.

Do not ask what the answer cost. Ask what it cost to produce work the business could use.