Metered Agents: When the Bill Depends on Which Model Answered

For three years, AI tooling was priced like software: a seat, a month, a predictable line item. That is over. Coverage of Cursor's 24 August 2026 change describes Auto moving off a single flat token rate to the routed model's list price, plus a $0.25-per-million-token processing rate on eligible third-party model requests for Teams and Enterprise — applied to input, output, and cached tokens, including bring-your-own-key traffic. Legacy enterprise Auto keeps the old flat rate only until 7 September 2026.[1]

Anthropic documents the other half of the picture for Claude Code: cost is token consumption, averaging about $13 per developer per active day and $150–250 per developer per month across enterprise deployments, with 90% of users staying under $30 per active day. Spend limits and workspace-level rate limits exist, which is an admission that the default is unbounded.[2]

Platform vendors are following. When Salesforce exposed its platform as agent-callable APIs and MCP tools, it also tied revenue to agent activity: customers pay for headless API consumption priced by usage, and contract separately for model inference. Agent traffic, not seat count, drives the meter.[3]

The operational consequence is specific. A long agent session pays input rates on a transcript that keeps growing, and under routed pricing the same loop can cost several times more on a day the router favors a frontier model. Budgeting a month of agent work now requires assumptions about routing behavior that the buyer often cannot observe.[1]

Independent survey data says control is lagging demand. ETR's 2026 spending research reports IT budget growth around 3.8% for the year, with 47% of organizations reporting AI spend moderately or substantially over plan, just 6% under, and 10% with no formal AI budget to measure against at all. When overruns hit, only 17% pause or scale back; the rest seek supplemental budget, absorb it, or reallocate.[4]

Where does the reallocation come from? According to the same research, external contractors and consultants first (61%), then legacy infrastructure modernization (49%), then non-AI software licenses (36%). Security is the notable exception — few organizations raid the security budget to fund AI overruns.[4]

A separate survey of technical leaders at mid-market and enterprise companies found 43% already over their AI budgets, most with annual AI budgets between $100K and $1M, and a stated preference for guardrails over behavior mandates: faced with a hypothetical 20% price increase, the most common responses were consolidating vendors and auditing usage with unclear ties to impact.[5]

Analyst estimates put worldwide AI spending in the trillions for 2026, and FinOps practitioners rank controlling token cost and usage in SaaS-delivered software as a top concern — largely because invoices arrive after the fact and rarely explain which workflow, team, or customer caused the charge.[6]

1) Measure cost per workflow, not cost per tool. A seat price tells you nothing about whether a recurring business output got cheaper. Attribute every paid request to a team, a workflow, and where possible a customer.

2) Set caps and spike alerts before the next invoice, not after. Detection in minutes beats reconciliation in weeks, and every major vendor now exposes some form of limit or admin API.

3) Budget for routing variance. If you cannot observe which model answered, treat the top of the routed price range as your planning number.

4) Bound the loops. Retry budgets, context-growth limits, and step ceilings are cost controls as much as reliability controls — an agent that retries indefinitely is a billing event, not a feature.

5) Include the invisible costs. Review time, exception handling, integration work, and the manual process still running in parallel during a pilot are part of the total, and they are usually what makes a promising workflow uneconomic.

6) Charge it back. Agencies and internal platform teams that cannot bill agent spend to the client or cost center that caused it end up subsidizing the enthusiastic.

7) Expect scrutiny on external spend. Consulting is the first line cut to cover an AI overrun, which is why work should be scoped short, tied to a number, and defensible inside the same quarter it is bought.

Vendor pricing moves quickly; the figures cited here carry dates for that reason, and negotiated enterprise agreements can differ from public list terms. Survey percentages describe responding populations, not every company. Where a pricing change is documented through industry coverage rather than a primary vendor page, we label it as coverage. Confirm current terms with your vendor before planning against them, and confirm your own baseline before quoting anyone's per-developer average — including ours.