Flat-Rate AI Pricing Is Dead. Anthropic Moves Agent Usage to Metered Credits June 15.

Starting June 15, Anthropic is splitting every Claude subscription into two buckets: interactive chat (your current plan limits) and programmatic usage (a separate monthly credit pool metered at full API rates). Agent SDK, claude -p, GitHub Actions, and third-party agents all move to the credit pool. Pro gets $20/month. Max 5x gets $100. Max 20x gets $200. Credits are per-user, don't pool, don't roll over.

This is the end of subsidized AI automation. And I think it was inevitable.

I run Claude Code in my personal projects every day, and I've watched my usage patterns change over the past year. Early on, I'd fire off a few prompts per session. Now I'm running multi-agent workflows with subagent spawning, MCP servers, and background tasks that burn through tokens I never consciously "use." Multiply that by every developer doing the same thing, and flat-rate pricing becomes a charity model.

As paddo.dev's analysis points out, the seat was never priced for the fleet. Flat-rate assumed human-speed interaction. Not 24/7 agent automation. This is the same trajectory as cloud compute: reserved instances gave way to spot pricing, which gave way to consumption-based billing. AI was always going to end up here.

What makes this move feel urgent is the convergence. Cursor's promotional pricing for Composer 2.5 ended May 25, and the Fast tier effectively doubled to $3/$15 per million tokens. Two major AI coding tools raising effective prices in the same week isn't coincidence. It's the market correcting.

What builders should do: Audit your programmatic Claude usage now. If you're running agent fleets, MCP-heavy workflows, or CI/CD integrations, your effective costs are about to change. The new /usage command in Claude Code v2.1.149 breaks down token consumption by skill, subagent, and MCP server. Use it. Build a baseline before June 15 so you're not surprised by the first credit bill. Start thinking about which agent tasks justify API-rate compute versus which ones you can batch or defer.