Anthropic is rolling back one of the most contentious developer policies of 2026. Starting June 15, paying Claude subscribers will once again be able to run programmatic workloads — including third-party agent harnesses such as OpenClaw — directly on their existing Pro and Max plans, but only inside a new capped credit pool that meters every token at API rates.
The change, reportedly first communicated to Max subscribers by email on May 13 and detailed across reports from VentureBeat, InfoWorld, and The New Stack, draws a clean line between interactive chat usage and what Anthropic now calls 'programmatic' usage. Anything driven by the Claude Agent SDK, the claude -p headless flag, Claude Code GitHub Actions, or a third-party harness built on the SDK will be billed against the new credit bucket.
How the credit pool works
The monthly allotment scales with the subscription tier and is designed to roughly match the plan's sticker price. Pro users receive $20 in Agent SDK credits per month, Max 5x users get $100, and Max 20x users get $200. Anthropic plans to email subscribers on June 8 so they can opt in, with the change going live a week later on June 15.
Two design choices have drawn the most attention from developers. First, credits do not roll over — unused balances expire at month-end. Second, once the bucket is empty, programmatic calls stop unless the account has 'extra usage' billing turned on, at which point overflow is charged at standard pay-as-you-go API rates rather than the bundled subscription rate.
A reversal of the April ban
The move is a notable climb-down from the policy Anthropic introduced in early April, which expressly prohibited Claude subscriptions from being used to power non-Anthropic agent frameworks. That decision pushed teams running automation pipelines onto direct API billing and forced harness projects like OpenClaw to either disable Claude support or route through customers' API keys.
By reinstating subscription-driven access — even with a meter attached — Anthropic is signaling that third-party agent ecosystems remain part of its commercial strategy, while still protecting margins on the most compute-intensive workloads. The Agent SDK credits sit alongside the company's other recent enterprise moves, including general availability of the Claude Platform on AWS and the rollout of Claude for Small Business.
What it means for builders
For independent developers, the practical effect is a smaller margin of safety. A Pro plan that previously absorbed bursts of automated work inside its general usage limits will now meter those bursts against a $20 envelope billed at API prices. Teams running production agent workflows will need to model their token consumption against the new caps and decide whether to enable overflow billing, upgrade to Max tiers, or move to direct API contracts.
For harness maintainers, the symbolic win matters: OpenClaw and similar projects are once again first-class citizens on paid Claude plans, just with a clearly drawn budget around them.
