US-based AI company Anthropic floated a plan to split "agent usage" out of Claude's regular paid-plan allowances and bill it through a separate monthly credit. The change was set to take effect on June 15, but Anthropic pulled it back within hours of going live and says current terms will stay in place for the time being. Developers had widely read the move as an effective price increase, which appears to be behind the reversal. Here we walk through what the shelved scheme would have done and the back-and-forth that has surrounded Claude's billing.

What the Withdrawn "Separate Billing" Would Have Done

At its core, the new scheme would have divided Claude usage into "interactive use" and "programmatic use," routing the latter through a dedicated credit pool.

Specifically, the developer-oriented Claude Agent SDK, the automation-focused "claude -p" command, Claude Code's GitHub Actions integration, and third-party apps built on the Agent SDK were all meant to draw from a new monthly "agent credit." Interactive use — chatting with Claude on the web, desktop, or mobile, running Claude Code in the terminal, and using the automation tool Claude Cowork — would have continued to count against normal subscription limits.

Until now, both programmatic and interactive use drew from the same plan allowance, so users never had to think about the distinction. The new scheme would have drawn that line clearly.

Per-Plan Credit Amounts and What Happens When You Run Out

The size of the credit was tied to each plan's price. For individuals, Claude Pro was set at 20 USD (about 3,200 yen), the higher Max 5x at 100 USD (about 16,000 yen), and the top Max 20x at 200 USD (about 32,000 yen). For teams, the standard plan was 20 USD per seat and the upper plan 100 USD per seat; for businesses, usage-based accounts were 20 USD and the higher enterprise plan 200 USD per seat.

*1 USD = 160 JPY

The credit was per individual account, with no sharing or pooling across teams and no rollover — unused credit would simply expire at the end of the billing cycle. If you used it up within the month, there were two options. Turn on overage and you could keep working without interruption, but that usage would be billed at pay-as-you-go API rates. Turn it off and your programmatic use was capped for the month until the credit refreshed in the next cycle. Developer-platform accounts that authenticate with an API key were excluded and would simply continue on pay-as-you-go billing.

A Shifting Stance on Third-Party Tools

Part of why the plan drew attention is that the treatment of third-party tools has changed repeatedly in a short span.

Not long ago, Anthropic switched to a policy that no longer let subscriptions cover usage of outside tools such as OpenClaw; at that point, users of external tools had to buy extra credit or accept pay-as-you-go API billing. The now-shelved scheme would have flipped that again, routing tools like Conductor and OpenClaw through the user's new agent credit, just like their own scripts — another turn in roughly a month.

Behind the churn lies the rapid spread of agent-style usage, led by Claude Code, and the resulting strain on compute resources. How to absorb that demand and allocate its cost is a line Anthropic is still trying to draw through trial and error.

For Now, Nothing Changes — What Users Should Watch

The change, which advanced right up to the eve of rollout, was ultimately withdrawn the same day. Anthropic says it will rework the agent-billing scheme for subscription users, and current terms hold for now. In other words, users do not need to reconfigure anything right away.

That said, the idea has not been scrapped — it could return in a different form. For developers who lean heavily on Claude Code and automation scripts in particular, a future revision would feed directly into their costs. Keeping an eye on Anthropic's next announcement is the safe move.

Summary

Anthropic put forward a scheme to carve Claude's agent usage out of regular allowances and bill it through a separate credit, only to withdraw it on the day it was due to take effect. The short-term wobble in its pricing reflects how hard it is for any provider to design the right charges as agent-style AI use spreads. How the reconsidered plan ultimately takes shape is a point worth watching for users and developers alike.