Metered AI Agents: Three Questions Before Your Next Renewal

Here's the short version of what happened. On May 13, Anthropic told subscribers that starting June 15, agents and automation running through a Claude subscription stop drawing from the flat monthly plan. They get a separate credit pool instead, $20 to $200 a month depending on tier, billed at standard API rates. No rollover, no pooling across a team. Credit runs out, agents stop, unless you've enabled pay-as-you-go at full API prices. Interactive chat is untouched.

The same workload was included in April. Then banned. Now metered. Three positions in three months, each announced by email. Analysts expect other vendors to follow, and GitHub Copilot is already moving toward credit-based billing.

If part of your AI plan involves agents doing work unattended, three questions are worth asking now, while it's a planning exercise instead of an invoice surprise.

1. Which of our AI usage is metered, and which is flat?

Most teams can't answer this today, because until recently the distinction didn't exist. Go tool by tool. Anything running on a schedule, in a pipeline, or through an API is now the category vendors are repricing. The interactive seats are not where the change is happening. The background work is.

2. What does the bill look like if usage doubles?

Under consumption pricing, a successful deployment is a more expensive one. The pilot that works gets used more, and the bill follows. Ask each vendor for the price at twice your current volume, in writing. If the answer involves a calculator and several assumptions, that uncertainty is itself the finding. Budget owners learned this with cloud compute, and an entire discipline, FinOps, exists because the contract price and the actual bill stopped matching.

3. Who can change the terms, and how fast?

The teams that built on third-party agent frameworks this spring did nothing wrong. The terms moved anyway, twice, with about 30 days' notice each time. Check your agreements for repricing windows and unilateral-change clauses. A workflow your auditors rely on shouldn't have a dependency on someone else's pricing page, and if it does, you want that documented as a known risk rather than discovered as one.

One structural note. A locally hosted platform, priced by capacity instead of consumption, doesn't have this exposure. The agents work all night and the bill doesn't move. That tradeoff isn't right for everyone, but if your data already can't leave your network, this month added a budget argument to the security one.