Starting June 15, Anthropic is splitting Claude subscription billing into two distinct pools: chat tokens for interactive use, and a separate "programmatic credit pool" specifically for agentic tool calling. The Pro tier at $20/month and Max tier at $200/month both include a monthly programmatic allocation that resets and does not roll over; additional tokens can be purchased beyond the included amounts. Anthropic frames the change as isolating expensive agent workloads from casual chat, citing the computational draw of "multiple rapid tool calls and high reasoning demand." Builder reaction in the announcement coverage is mixed; one developer characterizes it as "a downgrade dressed up as a feature."
The technical mechanism matters more than the framing. Until now, a single subscription quota covered both chat and any Claude session running tool calls in a loop — Claude Code, Cursor agents, or third-party wrappers like OpenClaw and Hermes that connect Claude as the planning brain for continuous tool execution. The new design moves loop-driven sessions onto a separate counter. For a team running automation at 24/7 cadence, the practical implication is that the unified bucket shrinks and a parallel bucket is sized for the agentic share. Whether that ends up cheaper or more expensive depends on the ratio between chat and agentic use the team was running — and Anthropic has not published the per-tier numerical limits at the time the news broke, only the dollar tiers.
Why this lands the way it lands: this is the second pricing-side change in the Claude Code era that builders are reading as a stealth quota reduction. The "5-hour window" rate-limiting episodes earlier in the year, and the August 2025 data-retention default flip, have made the user base wary of any change framed as "enabling agentic workloads at scale." The pattern that draws fire is bifurcation: separating what was unified into pools that each look smaller. Anthropic's argument is that agentic operations were always going to be priced separately at scale, and the dedicated pool makes that explicit rather than producing surprise overages. Both readings are defensible at the announcement stage; the real test is whether the published per-tier credit numbers leave room for the workloads people are already running.
For builders: budget for two pools instead of one going into June 15. Audit your Claude usage now by separating chat-style requests from agent-loop sessions — Claude Code, MCP server-driven workflows, anything that fires tool calls in a sustained loop — and check what the ratio looks like on a typical week. If the new programmatic allocation does not match that ratio, you have a month to either compress the agentic workload, plan for additional token purchases, or evaluate alternative providers for the loop-heavy paths. The full per-tier numerical limits are what to watch for in the next two weeks; the dollar tiers alone don't tell the whole story.