By ByteBot
Starting June 15, Anthropic is splitting Claude subscription usage into two separate meters. If you use claude -p, Claude Code GitHub Actions, or any third-party agent harness like OpenClaw, those calls will no longer draw from your standard subscription pool. They move to a dedicated monthly “Agent SDK credit” — $20 for Pro, $100 for Max 5x, $200 for Max 20x — billed at full API rates. When the credit runs out, your agents stop unless you’ve opted into extra usage and accepted open-ended charges. This resolves months of chaotic back-and-forth, and developers need to understand it before June 15.
Two Buckets, One Subscription
The clearest way to think about this: Anthropic is splitting your subscription into two pools that don’t share tokens. Interactive usage — Claude Code in your terminal or IDE, Claude.com, and Cowork — remains on your existing subscription limits, unchanged. Programmatic usage gets its own dedicated credit.
What counts as programmatic and moves to the new credit pool:
claude -p(non-interactive CLI invocations)- Claude Agent SDK calls
- Claude Code GitHub Actions
- Third-party apps and harnesses built on the Agent SDK (OpenClaw included)
Credits expire monthly and do not roll over. They’re priced at standard API rates — not the heavily subsidized rate implied by subscription pricing.
| Plan | Monthly Fee | Agent SDK Credit |
|---|---|---|
| Pro | $20 | $20 |
| Max 5x | $100 | $100 |
| Max 20x | $200 | $200 |
One thing worth noting: the Pro credit is exactly the same number as the subscription fee. Anthropic is pitching this as a bonus, but it’s the same $20 — now earmarked separately for programmatic use. Whether that reads as generous or cynical depends entirely on how much agent usage you were running before.
How We Got Here
This wasn’t a clean product decision — it was damage control dressed up as a feature. In April, Anthropic banned third-party harnesses outright. OpenClaw stopped working on subscription plans overnight. The stated reason was economics: agent workflows running “continuous reasoning loops” with large context windows were burning compute that a $20 subscription couldn’t cover. A single heavy agent session can consume 100,000–200,000 tokens. That’s a lot to subsidize at flat-rate pricing.
After community backlash and a brief, chaotic period where OpenClaw’s creator had his account temporarily suspended, Anthropic walked back the ban and landed on this metered middle ground: agents are back, but they now cost what they actually cost. VentureBeat reported on the reinstatement announcement framed specifically around the new metered structure.
The Extra Usage Decision
This is the most important operational decision you need to make before June 15. When your Agent SDK credit runs out, what happens next depends entirely on your extra usage setting.
Extra usage enabled: Programmatic calls continue at standard API rates. Your workflows keep running. But unless you’re actively monitoring token consumption, you could face a significant bill at month-end — especially with multi-step agent loops, large context windows, or high-volume GitHub Actions runs.
Extra usage disabled: When credits are exhausted, programmatic calls silently fail. Your automations break without warning. No surprise bills, but potentially broken CI/CD pipelines and stalled agents.
Neither is obviously the right choice. Teams with production agent pipelines probably need extra usage on — but they should also instrument their usage before June 15, not after. The Decoder’s coverage of the pricing mechanics is useful if you need the full breakdown.
The Industry Is Normalizing This
It’s worth stepping back. This is the second major AI coding tool this month to carve out separate billing for agentic use — GitHub Copilot made its own billing switch on June 1. AWS Kiro, which replaced Amazon Q Developer, has spec-driven pricing separate from IDE usage. The pattern is clear: AI vendors are pulling apart interactive and agentic usage into distinct billing lanes because the cost profiles are genuinely different.
Anthropic isn’t being punitive here. Subsidizing heavy agent workflows on a flat subscription was never going to work at scale. Axios reported the change as part of a broader competitive response to OpenAI’s Codex push — and that’s partly true. But the underlying driver is economics, not competition.
OpenAI is already moving: CEO Sam Altman announced two months of free Codex access for enterprises explicitly switching from Claude. The competitive pressure is real, but the metered billing trend is going to outlast any promotional offer.
What to Do Before June 15
Five concrete actions:
- Check your plan. Confirm whether you’re on Pro, Max 5x, or Max 20x and what credit you’ll receive.
- Watch for Anthropic’s email. Credits must be claimed. Anthropic will email subscribers by June 8 — don’t ignore it.
- Audit your programmatic usage. How many tokens per month do your agent calls,
claude -pscripts, and GitHub Actions consume? This determines whether your credit is sufficient or you’ll spill into API billing. - Decide on extra usage. Make a deliberate choice: hard ceiling (extra usage off) or continuous billing (extra usage on). Default to off unless your pipelines are critical.
- Re-evaluate your plan tier. If you’re on Pro and running meaningful agents, $20 in credit will likely run out quickly. The math might favor moving to Max — or switching to a direct API key for programmatic work and keeping the subscription for interactive use only.
The full policy details are in Anthropic’s official support article. XDA Developers has a clean breakdown of what changes for different user types.
