The weekly agent-spend audit that cuts the bill without hurting performance

The AI bill lands before the explanation. OpenAI lowered the standard API price for GPT-5.6 Sol to $4 per million input tokens and $20 per million output tokens, down from $5 and $30. For a small team paying for agents, the first question is which client, agent, and user produced the charge. If spend is attributed, reviewed weekly, and model choice is owned per task, the bill can fall without degrading performance. The audit turns a vendor invoice into an operating decision.

Price moves make routing a weekly problem

Agent spend looks like a utility line until someone asks who used it. A support agent, a sales assistant, and a research workflow can all call the same model and produce very different costs. The model sets the price; the task sets the usage. The same model can be cheap for one task and expensive for another.

OpenAI said its promotional pricing for GPT-5.6 Sol would remain available at least through November 21, 2026. Anthropic's Claude Sonnet 5 introductory price of $2 per million input tokens and $10 per million output tokens expires on August 31, 2026, with the standard price rising to $3 and $15 on September 1, 2026. Those moves change which model is cheapest for a given task. The dates matter because a model that is cheapest today can be more expensive next month. A team that does not re-check the routing will keep paying for last month's assumption.

Attribution turns spend into a control

Before caps, the team needs a view. Every request should carry enough metadata to answer which client, which agent, and which user. Without that, the dashboard shows total spend and hides the cause. The view should be simple enough to check in a weekly meeting. Separate paid work from experimental work. A client-facing agent has a different tolerance for waste than an internal experiment. If the team cannot tell the difference, the cap will punish the wrong thing.

Separate tokens from dollars in the view. Tokens explain usage. Dollars explain the bill. A cheap model can still become expensive if it is used on a long, repetitive task. A pricier model can be the right choice if it finishes in fewer turns. The distinction keeps the team from optimizing the wrong number.

Deterministic work is the easiest place to save. Classification, extraction, formatting, and routine drafting can often run on a cheaper model if the output still passes the team's checks. The test set acts as the guardrail. It should be small, current, and tied to the task, not to the model.

Department caps are the next control. A cap needs an owner with the power to stop non-urgent runs, reroute a task, or request a higher limit with a reason. The owner is accountable for the trade-off between cost and quality.

The weekly audit decides what changes

The checklist needs to be short enough to run in a standing meeting. It is the part of the week that turns spend into a decision. Keep the same dashboard and owner list in front of the team.

• Tag every request with client, agent, user, and task before it leaves the app.
• Sum tokens and dollars by client, agent, and user in a dashboard before the weekly review.
• Route deterministic work to the cheapest model that passes the task's test set.
• Set a hard weekly cap per department and stop non-urgent runs when it is hit.
• Name an owner for model choice per task and record the reason for each change.
• Re-run the prior week's spend after any price or model change.

End the meeting with a decision: whether a client is worth the spend, which agent is wasting tokens, and what stops when a department nears its cap. Name the next check. A client over the cap needs a pause, renegotiation, or a move of the work; a model change under test needs a stated result that justifies keeping it.