Claude Agent SDK Goes On A Separate Meter June 15. The $200 Monthly Credit Won't Cover A Day Of My Publisher Routines. Here's The Token Math.

Anthropic put Claude Agent SDK and claude -p on a separate meter starting June 15, 2026. The Max 20x plan ships with $200 in monthly Agent SDK credit. I run six publisher routines on Max 20x, and that $200 won't survive a full Tuesday.

What changed in three lines

Starting June 15, 2026, Agent SDK, claude -p, Claude Code GitHub Actions, and third-party apps built on the Agent SDK draw from a separate monthly credit pool instead of your subscription's regular usage allowance. The official help-center article spells out the scope and the failure mode if extra usage isn't enabled.

Credit by plan: Pro gets $20, Max 5x gets $100, Max 20x gets $200, Team Standard gets $20, Team Premium gets $100, and Enterprise allocates $200 Premium credit per seat. Credits reset monthly, don't roll over, and don't pool across teammates.

Interactive Claude Code in your terminal or IDE, Claude on web/desktop/mobile, and Claude Cowork all stay on the subscription pool. Only the unattended programmatic surfaces move. This is a price-discrimination split, not a perk, and the HN thread calling it a perk misses the architecture decision the split forces on everyone running real automation.

My six routines, in token math

Six publisher routines fire daily from claude.ai/code/routines on my Max 20x plan: publish-news, publish-dev, publish-design, publish-marketing, publish-founders, publish-business. A seventh, publish-ai, produced this article. Each routine wakes a Claude Code session, driven by claude -p, against a Cloudflare admin worker that owns the editorial brief and SERP scrape, then handles the publish handoff.

Per-fire token cost on Opus 4.7 ($5 input / $25 output per million) lands around $5–$15 depending on how much SERP context the brief pulled in. On a busy publishing day I'm running 20–30 fires across the seven routines. That's $100–$450 in token value per day at headline API rates.

The $200 monthly Max 20x Agent SDK credit covers roughly half a day to two days of that cadence. Beyond the credit, additional Agent SDK requests bill at standard API rates if extra usage is enabled in the account. If it isn't, the pipeline halts silently until the next billing cycle. No warning email mid-month. No graceful degradation or pause-and-confirm. Requests just stop.

This is the workload Anthropic's Boris Cherny was describing when he called the subscription-driven third-party tooling pattern "really hard to do sustainably." A $200 subscription was quietly consuming hundreds to thousands of dollars in token value, day after day, on uncached agent contexts that the subscription was never priced to subsidize. Greyhound Research's Sanchit Vir Gogia framed the broader shift correctly: the "all-you-can-eat era" isn't ending; it's moving onto a separate menu. The split is the menu.

I wrote about the +50% weekly limit bump on Claude Code a few weeks ago. That bump was an apology. This split is the actual policy.

What migrates, what stays, what dies

Migrates to a direct API key. Any unattended workload where the per-token rate is the binding constraint, not the convenience of "I'm signed into Claude." Publisher routines, scheduled ops audits, CI/CD scripts, Claude Code GitHub Actions that fire on every PR. The cost model doesn't change when you move; you just stop pretending the subscription was carrying it. The honest move is to wire these through AI Gateway with BYOK so you keep observability and rate limiting in one place.

Stays on the Max plan. Interactive Claude Code sessions in the terminal, pair-coding sessions, exploratory builds, long-context human-in-the-loop work. This is what the subscription was designed for, and the unmetered ceiling on this kind of usage is still the best deal in the industry for individual operators. My P&L argument for staying on Max despite the Codex two-month-free offer hasn't changed; the math gets sharper after June 15, not weaker.

Dies on June 15. Every "I built a side-project agent that quietly hits Claude all day from a $20 Pro plan" stack. The moment automation gets useful enough to fire regularly, you're paying API rates. The single-developer hobby agent economy that lived inside subscription limits is over.

One detail the support article doesn't flag clearly: Claude Code GitHub Actions integrations move into the metered pool even when the action is "interactive" in spirit. A review bot that comments on every PR feels like a tool, but it runs on claude -p under the hood, and it's now on the meter. Audit your action workflows this week.

The operator playbook this week

Open the Anthropic console and check whether extra usage is enabled. If it isn't, your pipeline halts silently the moment your credit drains. The failure mode: no alert, no email, no degraded mode. Just silence. For a publisher cadence like mine, silence on a Tuesday morning is a real outage.

For every agent or script you run on claude -p, log token usage per invocation now. Not in June. You cannot budget what you cannot measure, and the credit drain is per-user, so a Team admin can't see aggregate burn the way they could before. Wire a usage tracker into your worker before the June 8 prep email lands.

Set prompt caching aggressively on Agent SDK system prompts. Caching survives the meter switch. The cache-miss penalty on uncached third-party agent contexts is the structural cost driver Anthropic priced this split against, and it's the lever you control.

Architect the workload boundary explicitly. One worker calls Anthropic via your API key for unattended fires. The subscription stays clean for interactive Claude Code. Do not commingle. The credit-first drain order means the cheaper subscription-equivalent usage exhausts your credit first, leaving the genuinely expensive workloads to bill at API rates anyway. You lose the discount on the wrong side of the workload.

If you administer a Team or Enterprise account, do the per-user audit before June 8. "Each eligible user on your team claims their own credit" means the single-admin spend visibility you had before is gone, and the per-seat $200 on Enterprise Premium can hide a six-figure annualized burn across 40 seats.

Why this happened, and what's next

The April 2026 OpenClaw cut was the warning shot. Anthropic stopped covering third-party Agent SDK frameworks under Claude subscriptions and offered credit bundles to ease the transition. Most operators read it as a one-off vendor dispute. It wasn't. It was the policy direction in miniature.

The June 15 split generalizes that policy across every programmatic surface Anthropic owns. Every external automation entry point is now metered. Notion's August external-agent credit cliff is the same pattern from the platform side: metered economics for any agent that isn't the platform's own.

Gogia named the broader move correctly. GitHub is migrating Copilot to a credit system that "closely resembles Anthropic's latest changes." OpenAI has lived on API-style pricing since GPT-4. Over the next 12–24 months, every major lab will carve out separate consumption pools for agents, premium models, tool use, and background tasks. The unified subscription that covered everything was a 2024 artifact, not a permanent feature.

The honest operator move: stop pricing automation against subscription tiers. Start pricing it against the underlying token cost. The era of treating Max 20x as cheap CI/CD compute ends on June 15, and the operators who model that change into their pipeline this month will be the ones whose Tuesdays survive.

Will my Claude Code sessions still work the same after June 15?

Do unused Agent SDK credits roll over each month?

What happens when my monthly Agent SDK credit runs out?

Should I migrate my agent workload to a direct API key?