Anthropic just figured out how to stop AI power users from draining their entire subscription business model — by inventing a two-tier credit system that makes "unlimited" mean two different things at once.
The Summary
- Anthropic reversed its April ban on using Claude subscriptions for third-party agents like OpenClaw, but split subscription credits into two pools: general use and "Agent SDK" credits
- Users were burning hundreds or thousands of dollars in compute while paying $20-$200/month subscriptions, creating an unsustainable arbitrage
- The new system gives programmatic agent access back, but caps it with fixed credits instead of letting power users feast on the general subscription pool
The Signal
Anthropic's reversal isn't magnanimous. It's economic triage. In April, they realized their subscription pricing had a fatal flaw: heavy OpenClaw users could turn a $200/month Claude Max subscription into thousands of dollars of compute. The company explicitly banned third-party agents because autonomous tools running 24/7 consumed compute like a data center, not like a human typing queries.
Now they're back with Agent SDK credits, a separate bucket for programmatic use. You still get your subscription. You still get OpenClaw access. But the all-you-can-eat buffet is closed. Agent credits are capped, metered, finite.
"Programmatic usage is no longer subsidized by the general subscription pool but is instead restricted to a fixed, new allocation."
This is the first major pricing adjustment that acknowledges autonomous agents aren't just power users. They're a different species of customer. Human users open Claude, ask questions, close the tab. Agents loop. They iterate. They run workflows while you sleep. The usage pattern doesn't fit flat-rate pricing.
Here's what Anthropic learned the hard way:
- Subscription models assume human behavior: sporadic, bursty, bounded by attention and time
- Agent behavior is continuous, optimization-driven, and scales with automation, not biology
- Mixing both in one pricing tier creates arbitrage opportunities that bankrupt your margin
Other AI companies are watching. OpenAI has ChatGPT Pro at $200/month with "unlimited" GPT-4. Google offers similar tiers. If agents go mainstream, every AI subscription will face the same breaking point. Do you cap usage? Split credit pools? Force agent users to API pricing?
Anthropic chose the split. It's elegant but awkward. Subscribers now have two wallets: one for chatting with Claude, one for letting Claude build things autonomously. The implication is clear: using AI as a co-pilot is cheap. Using AI as a worker is expensive.
The Implication
If you're building agent tools or workflows, price like Anthropic just did. Assume your power users will consume 10-100x the average. Don't subsidize automation with conversational pricing. If you're a Claude subscriber running agents, budget for the reality that Agent SDK credits will run out faster than you think. And if you're watching this space, understand what just happened: AI companies are learning that agents aren't just features. They're a business model fork.
The companies that figure out agent pricing first will own the automation economy. The ones that don't will keep banning their best users.