Six weeks ago, Anthropic did something that quietly disrupted the workflows of thousands of small business owners: it cut off the ability to use Claude subscriptions with third-party AI agents. Tools that had become central to automating customer service, research, scheduling, and internal processes suddenly required users to either pay for expensive API access or abandon their setups entirely.
Today, Anthropic reversed that decision - with conditions that every small business owner should understand before assuming things are back to normal.
What Happened in April
On April 4, 2026, Anthropic blocked Claude subscribers from using services like OpenClaw - the open-source AI agent platform that lets businesses run Claude-powered automation through tools like Discord, Telegram, and custom workflows. The reason was financial, not technical: subscribers paying $20 to $200 per month were running agents that consumed thousands of dollars worth of compute, with Anthropic absorbing the difference.
The problem was real. Agentic workflows, by design, process information repeatedly. Unlike Anthropic's own products, which are engineered to maximize "prompt cache hit rates" (essentially reusing processed data to cut costs), third-party agents often bypassed those efficiencies entirely - burning through tokens at rates that made flat-rate subscriptions unsustainable.
What Changed This Week
Anthropic announced today via its developer account on X that it is creating a new category of credits specifically for programmatic and agent use - called "Agent SDK credits." These are separate from your existing Claude subscription usage and are billed at API rates rather than the flat-rate model.
Here is how the credit amounts break down by plan:
- Claude Pro ($20/month): $20 in Agent SDK credits
- Max 5x ($100/month): $100 in Agent SDK credits
- Max 20x ($200/month): $200 in Agent SDK credits
- Team plan: $100 per seat
- Enterprise: $200 per seat
The credits are dedicated, non-rollover, and cannot overflow into your general subscription usage. If you burn through them in two weeks of agent activity, you stop until the month resets - or you buy additional usage credits at API rates.
What This Means If You Run AI Automation
The practical implications depend on how you have been using AI agents in your business.
If you run light automation - a bot that handles a few dozen customer inquiries a day, or a scheduled task that pulls data and summarizes it - the new credit allocation is probably sufficient. A $20 agent credit at API rates covers a meaningful amount of work for low-volume use cases.
If you run heavier automation - agents that process documents, respond to high volumes of messages, or run continuously throughout the day - you will likely hit the cap. The math has changed: what used to feel "included" in your subscription is now metered, and the meter runs fast under real workloads.
If you abandoned AI agents after the April ban - this is your signal to re-evaluate. The access is real. But build your workflows with the credit cap in mind from the start, rather than discovering the limit when your agent goes quiet mid-month.
The Bigger Picture
This policy shift reflects a tension that is going to become more common as AI adoption deepens: flat-rate pricing built around human usage patterns is fundamentally incompatible with autonomous agents that can run 24 hours a day. Anthropic tried a blunt solution (full prohibition), found it created too much friction, and landed on a metered compromise.
For small business owners, the lesson is that the economics of AI automation are still being worked out in real time. The tools are becoming indispensable faster than pricing models can keep up. Anthropic's new system is a reasonable middle ground - but "reasonable" assumes your workload fits inside the credit box. Know your usage before you rebuild your workflows around it.
Terry Blake explains how technology and regulation work for small business owners without the jargon. Sources: VentureBeat, Anthropic Claude support documentation.