Claude Agent SDK Billing Split: What the June 15 Change Breaks and How to Fix It

Two days ago — June 15, 2026 — Anthropic flipped a switch that affects every developer running automated Claude pipelines. Headless invocations, GitHub Actions automations, and anything driven by the Claude Agent SDK stopped drawing from your subscription quota and started drawing from a separate monthly credit pool. When that pool empties, your automation stops — full stop, no graceful degradation — unless you've enabled overflow billing.

If you logged in this week and found a depleted credit balance you didn't expect, or if you're building pipelines and haven't updated your billing config yet, this is the post you need.

What changed on June 15

Anthropic announced on May 14 that they'd be splitting Claude subscription billing into two distinct pools, effective June 15. The reason, stated plainly: interactive use and autonomous agent use have very different cost profiles. A human in the loop naturally throttles usage. An unattended pipeline doesn't.

Before June 15, running claude -p "fix the CI failure" in a GitHub Action counted against the same quota as opening Claude.ai and asking a question. After June 15, it draws from a separate monthly credit budget instead. The subscription quota is unchanged. The credit pool is new.

The two billing buckets

Your Claude subscription now has two independent spending pools. Bucket one is unchanged: interactive Claude use — chat, Claude Code in the terminal when you're the one driving, Claude Cowork collaboration — all draw from your plan's existing limits. Bucket two is new: the agent credit pool.

Credits are consumed at standard API rates — the same per-token prices you'd pay if you called the API directly. No markup, no subscription discount applied. Unused credits do not roll over to the next month.

What's affected (and what isn't)

The split is clean: autonomous and programmatic usage moves to the credit pool; human-in-the-loop usage stays on the subscription. The clearest signal is whether a human is actively typing.

Moved to the agent credit pool

• claude -p "..." — the headless print flag, the backbone of most CI/CD and cron automations
• Claude Code GitHub Actions — any workflow step that invokes Claude Code
• Direct Claude Agent SDK calls (@anthropic-ai/claude-agent-sdk /claude-agent-sdk)
• Third-party apps that authenticate through the Agent SDK on your behalf
• Scheduled /loop tasks running headlessly

Unchanged — stays on subscription

• Interactive Claude.ai chat sessions
• Claude Code in the terminal when you are actively driving the session
• Claude Cowork collaborative sessions
• Claude Desktop app usage

The SDK rename and migration

The billing split arrived alongside a package rename. In September 2025, the Claude Code SDK was renamed to the Claude Agent SDK to signal that it isn't only for coding agents. If you were on the old package names, June 15 is the forcing function to migrate.

# Remove the old package
npm uninstall @anthropic-ai/claude-code

# Install the new one
npm install @anthropic-ai/claude-agent-sdk

# Remove the old package
pip uninstall claude-code-sdk

# Install the new one
pip install claude-agent-sdk

Python users have one breaking type change to fix after the swap. The options type was renamed:

# Before (claude-code-sdk)
from claude_code_sdk import ClaudeCodeOptions

options = ClaudeCodeOptions(
    system_prompt="You are a code reviewer.",
    max_turns=10,
)

# After (claude-agent-sdk)
from claude_agent_sdk import ClaudeAgentOptions

options = ClaudeAgentOptions(
    system_prompt="You are a code reviewer.",
    max_turns=10,
)

TypeScript users get a smoother ride: the package name changes but the exported types are identical. A find-and-replace on the import path is sufficient.

// Before
import { query, type ClaudeCodeOptions } from "@anthropic-ai/claude-code";

// After
import { query, type ClaudeAgentOptions } from "@anthropic-ai/claude-agent-sdk";

Estimating your monthly credit spend

The key question is whether your automation fits inside the credit pool or spills over into overflow billing. The math isn't complicated once you have usage data.

// Approximate token costs (as of June 2026 standard API rates)
// Opus 4.8:  $15 / 1M input tokens,  $75 / 1M output tokens
// Sonnet 4.6: $3 / 1M input tokens,  $15 / 1M output tokens
// Haiku 4.5:  $0.8 / 1M input tokens, $4 / 1M output tokens

// Example: a nightly code-review pipeline
const runsPerMonth = 30;           // once per day
const inputTokensPerRun = 50_000;  // ~40KB diff + system prompt
const outputTokensPerRun = 2_000;  // summary + inline comments
const model = "claude-sonnet-4-6"; // Sonnet 4.6

const inputCost = (inputTokensPerRun / 1_000_000) * 3;   // $0.15 per run
const outputCost = (outputTokensPerRun / 1_000_000) * 15; // $0.03 per run
const costPerRun = inputCost + outputCost;                 // $0.18

const monthlyCost = costPerRun * runsPerMonth; // ~$5.40 / month
// Fits comfortably in a Pro ($20) credit pool.

The most dangerous pattern is fan-out automations: scripts that spawn many parallel agents, each with a large context. A single orchestrator that launches 10 subagents at 100K tokens each will cost roughly $3–5 per invocation using Sonnet. Run that hourly and you'll hit a $200 credit pool in under three days.

Checklist: what to do right now

If you haven't acted since June 15, work through this list today. Most items take under five minutes.

## Claude Agent SDK Billing Migration — June 2026

### Immediate (do today)
- [ ] Open billing dashboard → confirm agent credit pool balance
- [ ] Enable overflow billing if any automation is business-critical
- [ ] Search repo for "@anthropic-ai/claude-code" → update to "@anthropic-ai/claude-agent-sdk"
- [ ] Search repo for "claude-code-sdk" (Python) → update to "claude-agent-sdk"
- [ ] Search Python code for "ClaudeCodeOptions" → rename to "ClaudeAgentOptions"

### This week
- [ ] Audit every cron / CI step that uses `claude -p`
- [ ] Estimate monthly token cost per pipeline (see formula above)
- [ ] Set billing alerts at 50%, 80%, 100% of credit pool
- [ ] Decide which automations to route to a smaller model (Haiku 4.5) to stretch credits

### Optional but recommended
- [ ] Add --model flag explicitly in all claude -p invocations (don't rely on defaults)
- [ ] Install skills-hub ops/claude-agent-billing-audit to run automated audit
- [ ] Pin SDK versions in package.json / requirements.txt after migration

The overflow billing toggle lives in Account Settings → Billing → Agent Credits. It's off by default. With it on, usage beyond the monthly pool bills at standard API rates with no hard stop. With it off, automation halts and returns an error when credits hit zero.

The billing split isn't a penalty — it's the infrastructure growing up around a real use pattern. Autonomous agent runs are fundamentally different from interactive sessions, and separate billing makes that visible. The teams that handle this well will audit their pipelines, right-size their model choices, and come out with cleaner cost attribution than they had before.

If you want an automated audit of your current setup, the ops/claude-agent-billing-audit skill on skills-hub.ai will scan your repo for affected invocations, estimate monthly token cost, and produce a migration checklist tailored to your codebase.

npx @skills-hub-ai/cli install claude-agent-billing-audit