【6/15〜】Claude Code の課金変更、何が対象で何が対象外かを公式情報だけで整理 Starting June 15, 2026, programmatic usage of Claude via the Agent SDK, claude -p, and Git…

Anthropic is making a notable change to its Claude Pro and Max subscription plans starting June 15, 2026. Programmatic access to Claude — including the Claude Agent SDK, the claude -p command-line interface, and GitHub Actions integrations — will no longer count against the standard subscription usage allowance. Instead, this class of usage will be billed separately, effectively splitting how interactive and automated workloads are metered.

The distinction Anthropic appears to be drawing is between hands-on, interactive use of Claude Code and automated or batch-style invocations that resemble API calls. Manual, conversational sessions are expected to remain within the subscription tier, while non-interactive calls triggered by scripts, CI pipelines, or external orchestration layers will fall under the new billing category. The boundary isn't always obvious, and developers are advised to consult the official documentation carefully as it is updated.

The motivation behind this shift is likely tied to the explosion in agentic use cases that followed the launch of Claude Code. What started as a developer productivity tool has quickly been embedded into enterprise CI/CD workflows, batch summarization pipelines, and multi-step agent chains that can consume enormous token volumes — usage patterns that are structurally very different from a single user chatting with an assistant. Keeping both under one flat-rate umbrella appears to have become financially unsustainable for Anthropic.

This trend isn't unique to Anthropic. OpenAI and Google have each introduced or explored tiered and consumption-based billing for automated and agent-driven workloads, signaling a broader industry shift toward separating conversational quotas from programmatic throughput. As AI agents become first-class citizens in software development pipelines, per-token or per-execution pricing models are likely to become the norm rather than the exception.

For developers with existing automation built on Claude Code, the practical implication is clear: any GitHub Actions job, cron-triggered script, or autonomous agent loop that calls Claude programmatically could generate additional charges after June 15. Setting spend limits and auditing current usage before the cutover date is a prudent step. Anthropic has indicated it will continue updating its documentation to clarify how it distinguishes between subscription-covered and separately billed usage, but the onus remains on developers to track those updates and adjust their integrations accordingly.

As agentic AI becomes more deeply woven into software development and business automation, billing models will almost certainly grow more granular. Staying informed about these policy changes is no longer just an accounting matter — it directly affects how teams architect their AI-assisted workflows and manage infrastructure budgets.