RIP token maxxing: the subsidy on programmatic Claude usage is ending

RIP token maxxing. Mid-June, the slushie machine turns off - unless you’re willing to spend real money to keep it running.

• The GitHub Action that fires Claude on every PR? Metered, unless you spend $$$.
• The claude -p script running in cron at 2am? Metered, unless you spend $$$.
• The third-party app you built on the Agent SDK and quietly pointed at your Pro plan? Metered, unless you spend $$$.

Seeing the pattern yet?

What’s changing

Starting June 15, programmatic Claude usage gets its own meter. Agent SDK calls, claude -p, GitHub Actions, anything built on the SDK - all of it now draws from a new monthly credit, separate from your interactive session limits. The credit equals the price of your plan: a $20 plan gets a $20 credit, a $100 plan gets a $100 credit, a $200 plan gets a $200 credit.

Anthropic’s framing is “claim a dedicated monthly credit,” which is one way to describe it. Another way: you now get exactly what you pay for, and not a cent more.

The arbitrage is what’s dying

We’d quietly built a whole generation of agentic workflows on top of an implicit subsidy. Nightly audit loops, doc generators, PR reviewers, custom agents pointed at the SDK - all burning hundreds of dollars a month of real inference cost for the price of a single Pro plan seat. That gap between what those workflows actually cost to run and what anyone was paying for them was never intentional; it was just a leak nobody had gotten around to patching.

They’ve patched it now. Which means every prompt has a visible dollar cost going forward. Every “send it and see what happens” run is money out the door. Every agent you spawn draws directly on your meter, in a way you can no longer avoid noticing.

The question that actually matters now

For a lot of teams, the operating question has quietly been “how much can I get out of my subscription” - treat the plan as a flat fee, run as much through it as the interactive limits will allow, and don’t think too hard about what any individual prompt costs. That question stops being answerable, because the flat-fee illusion is gone for anything programmatic.

The question that replaces it is less comfortable: is my team actually getting value out of every prompt they fire? We’ve spent a couple of years training a generation of developers to treat AI usage like a slot machine - pull the lever, see what comes out, pull again if it wasn’t great. That instinct is about to get expensive in a way it wasn’t before, because now there’s a running total attached to every pull, and it’s visible in real time instead of buried in a subscription line item.

If you can’t currently tell me which of the prompts your team ran this week were actually worth the money - not which ones ran, which ones were worth it - that’s the gap you need to close before June 15, not after.