Anthropic announced a billing split in May separating agent workloads from subscription pools into metered credits. The change was paused before its June 15 go-live, but the logic is visible elsewhere: GitHub Copilot shifted to usage-based AI Credits on June 1, and one agentic coding session now costs $30–40.
Anthropic's proposed credits were per-user, non-pooled, stop-on-empty. Teams would need to classify workloads, attribute costs to specific accounts, and pre-authorize agent consumption rather than discovering it post-hoc. Flat-rate subscriptions let organizations defer all of those questions indefinitely. A meter makes deferral uncomfortable.
Governance infrastructure asks whether an agent acted within scope. Billing forces something more granular: what counts as a unit of work, who pays for it, what happens when the budget runs out. That's specification work, and pricing imposes it whether organizations are ready or not. The competitive scramble that followed suggests nobody wants to force that clarity on customers yet. But an estimated 15–30x subsidy gap between subscription price and API-equivalent compute means the deferral has an expiration date.
Affected workloads: Agent SDK, claude -p, Claude Code GitHub Actions, third-party agents. Interactive chat unchanged.
Credit tiers (proposed): $20/mo Pro, $100/mo Max 5x, $200/mo Max 20x. Billed at API list rates, no rollover.
The subsidy in dollars: Heavy users accessing an estimated $300–600 of API-equivalent compute on a $20 plan.
Copilot reaction data: 10x–100x cost swings on heavy workloads; one editor burned 82% of monthly credits on day one. GitHub Community thread: 958 downvotes, 24 upvotes.
Competitive moves: OpenAI offered two months free Codex Enterprise. Anthropic raised Claude Code limits 50% through July 13.
Timeline: May 14 announced → June 15 paused.