Anthropic Reverses Its Third-Party Agent Ban, But the Permission Slip Costs More Than It Looks
Anthropic's $20 Comeback Gift for Third-Party Agents: Enough Runtime for About an Hour a Day
Anthropic is letting third-party agents back into Claude subscriptions — with a catch that makes the catch the story. Starting June 15, Pro users get $20 per month in Agent SDK credits; Max 5x gets $100; Max 20x gets $200. Credits do not roll over. When they are gone, programmatic usage stops unless pay-as-you-go billing is enabled. Developer reports cited by VentureBeat show a single OpenClaw agent running for one day can burn $1,000 to $5,000 in API costs. The flat-rate subscription that once let a solo developer run agents overnight now caps out before the workday begins.
The credit structure is the first concrete detail Anthropic has released since reversing its April 4 ban on third-party agent access. The reversal — restoring the ability for Claude Pro and Max subscribers to hook their subscriptions to OpenClaw and similar tools — was announced May 13 via @ClaudeDevs. But the mechanism of that restoration was not specified until now. Developer Theo Browne put the cut plainly: usage for external tool users is reduced by 25x. "They are disguising this as free credits," Browne posted. "Do not fall for it."
Peter Steinberger, the creator of OpenClaw, saw the pattern earlier. "Funny how timings match up," he posted. "First they copy some popular features into their closed harness, then they lock out open source." The reference is to Anthropic's replication of agent patterns from the OpenClaw ecosystem into Claude Code, followed by the April attempt to restrict programmatic access. The credit system is the sequel: the lockout is now priced.
The math behind the ceiling is concrete. Anthropic's Claude 3.5 Sonnet API runs approximately $3 per million input tokens and $15 per million output tokens at standard rates. A typical OpenClaw agent task — multiple tool calls, browser interactions, iterative reasoning — consumes tens of thousands of tokens per minute. At a conservative 100,000 tokens per session minute, $20 of API credit covers roughly 30 to 60 minutes of continuous agent runtime. At the usage intensities OpenClaw reaches during autonomous task loops, the budget is gone in under an hour.
The technical reason for the original ban was cache efficiency. Third-party agents like OpenClaw bypass the prompt cache hit rates that Anthropic's own Claude Code and Claude Cowork are engineered to maximize. Boris Cherny, Anthropic's Head of Claude Code, said third-party services were "not optimized" and "really hard for us to do sustainably." The new credit system does not fix that underlying inefficiency. It prices it. Anthropic is no longer absorbing the cost of unoptimized agent code; subscribers are.
The reversal is arriving at an awkward moment for Anthropic's competitive position. Kun Chen, a former engineer at Meta, Microsoft, and Atlassian who built agentic workflows on Claude, announced he was moving to OpenAI after the April ban. "Anthropic's only lead was on coding, and GPT 5.5 has flipped that already," Chen posted. The credit system does not address the talent migration; metered access reinforces the signal that Claude is not the place to run production agents.
Anthropic is framing the credit as additive value. Lydia Hallie, an Anthropic technical staff member, posted that subscribers pay "the same subscription, same price per month" and the credits are extra. That framing is technically accurate but economically misleading for anyone who used their subscription to run agents. The $20 Pro credit is not a bonus; it is a hard ceiling on a use case that previously had none.
The credit structure does offer something the ban did not: a path for OpenClaw and similar tools to authenticate legitimately via the Agent SDK rather than routing through the API at full pay-as-you-go rates. For light, intermittent use, the credits are not zero. A developer who runs a few agent tasks a month might find the Pro credit sufficient. For anyone running agents as infrastructure, the credits are a floor, not a ceiling, and the floor is low.
What Anthropic has done is convert a product problem into a pricing tier. The unsustainable compute cost of third-party agentic workflows is real; Anthropic's own infrastructure constraints are real. But the solution is not a $20 credit that expires at the end of the month. It is either a genuinely competitive API product for agentic workloads, or a tier that reflects what autonomous agents actually cost to run.
The April ban pushed power users like Chen to OpenAI. The credit system keeps them there. For the developer still on the fence about where to run their agents, the math is simple: $20 buys one hour of OpenClaw. Anthropic's subscription no longer includes the rest.