If you've been running Agent SDK pipelines on a Claude Pro or Max subscription, you've been getting a deal that was never meant to last. Anthropic formalized the end around May 14. June 15 is the date it closes.

Starting then, usage of the Agent SDK, the claude -p command, Claude Code GitHub Actions, and third-party apps that authenticate via the Agent SDK moves to a separate monthly credit pool billed at standard API list prices. Pro subscribers get $20/month in credits; Max 5× gets $100; Max 20× gets $200. Credits expire at the end of your billing cycle with no rollover. Hit the ceiling and all automated requests hard-stop — unless you've manually enabled overflow billing first.

Interactive usage is explicitly unaffected. Chatting with Claude on the web, desktop, or mobile, running Claude Code interactively in the terminal, and using Claude Cowork all continue drawing from the existing subscription pool.

This is Anthropic's third attempt this year to close the same structural gap. They banned subscription-level agent usage in February. Tightened the ban in April. The bans broke too many legitimate use cases. The metered approach is more surgical: the usage can happen, it just costs what it actually costs.

The backstory involves a pattern sometimes called "subscription arbitrage." Users paying $20 to $200 per month for a Claude subscription used frameworks like OpenClaw to run multi-hour autonomous tasks, consuming tokens that would cost hundreds or thousands of dollars at API prices. The gap between subscription cost and actual API cost for identical agent workloads has been measured at roughly 15–30× at the median. On the heaviest long-context workloads — the kind a detailed independent analysis calculates at 12×–175× depending on session length and model — the gap was never going to survive indefinitely.

Source spread

• The New Stack — Anthropic splits billing again [builder] — clean technical breakdown of what moves pools and what doesn't.
• TechTimes — Credit pool replaces flat-rate access [builder] — covers the economics and the OpenClaw arbitrage backstory in depth.
• VentureBeat — Anthropic reinstates OpenClaw with a catch [skeptic] — flags the cost impact for third-party tool users who didn't realize they were being subsidized.
• GitHub Gist — Full workload math analysis [builder] — independent detailed math on the 12×–175× effective price increase range, with workload breakdowns.

Pros & cons

What's reasonable:

• The economics were never real. $20/month flat doesn't cover multi-hour Opus 4.8 sessions. Anthropic was absorbing losses to keep price parity — fine as a launch subsidy, not as an indefinite operating model.
• Overflow billing can be enabled. For builders who need uninterrupted automated access and are willing to pay API prices, there's no hard ceiling. You just have to opt in explicitly.
• Interactive Claude Code in the terminal is carved out. If your Claude Code usage is developer-interactive rather than CI/CD-automated, nothing changes for you.
• The credit tiers scale with the subscription. Max 20× users get $200 in credits — ten times what Pro users get, which is internally consistent.

What warrants watching:

• $200/month at Opus 4.8 API prices ($5/M input, $25/M output) doesn't last long in heavy production agent workflows. Forty million input tokens sounds like headroom until you model a few runs with large codebases.
• No rollover. Bursty workloads — end-of-sprint batch jobs, weekly autonomous reports — can't bank unused credits. Use them or lose them.
• Third-party apps using Agent SDK authentication are in scope. Cursor extensions, OpenClaw users, and others may not know this billing split is coming until their automated workflows stop on June 15. Anthropic says it's emailing account holders; that's not the same as the downstream apps providing a clear heads-up.
• The credits must be claimed before June 15. Anthropic is sending instructions to account emails. If you manage multiple Claude accounts across multiple projects, each one needs attention separately.

Further reading

• The New Stack — Anthropic splits billing again: Agent SDK gets separate credit pools
• TechTimes — Anthropic ends subscription subsidy for agents
• VentureBeat — Anthropic reinstates OpenClaw and third-party agent usage with a catch
• GitHub Gist — Canonical workload math and effective price increase analysis
• CodersEra — What every Claude Code and Agent SDK user must do