Anthropic has announced Claude Agent SDK credits, enabling third-party autonomous AI agent harnesses such as OpenClaw to run again.



Anthropic has announced the introduction of a monthly 'Agent SDK Credit' tier to Claude's paid plans, separating the use of the Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party apps authenticated with the Agent SDK from the regular subscription tier. As a result, Pro, Max, Team, and Enterprise users will receive dedicated credits worth between $20 (approximately 3,100 yen) and $200 (approximately 31,000 yen) depending on their plan, while regular use of Claude chat and interactive Claude Code will continue to be subject to the existing subscription limits.

Using the Claude Agent SDK with your Claude plan | Claude Help Center
https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan

Agent SDK credits are a monthly, dedicated usage allowance specifically for calling Claude from programs. They apply to using the Claude Agent SDK in your own projects, the non-interactive mode of Claude Code (claude -p), Claude Code GitHub Actions, and third-party apps that authenticate to a Claude subscription via the Agent SDK. They do not apply to Claude conversations on the web, desktop, or mobile, or to interactive Claude Code or Claude Cowork.



Agent SDK credits are granted on an individual basis and cannot be shared or combined within a team. Unused credits are not carried over to the next month. Agent SDK usage is first consumed from monthly credits. After the credits are used up, you will be charged at the standard API rate only if additional usage is enabled. If additional usage is not enabled, Agent SDK requests will be suspended until the next renewal.

Technology media outlet VentureBeat has described this measure as a policy shift by Anthropic.

In April 2026, Anthropic restricted the use of Claude subscriptions with third-party agent tools such as OpenClaw, but this new system allows external agents like OpenClaw to be used via subscriptions once again.



However, Anthropic's initial decision to block the service stemmed from a mismatch between the design of its subscription service and the actual usage of its agents. VentureBeat reports that while Claude Pro and Max users were paying monthly fees ranging from $20 to $200, they were consuming tokens that could have been worth hundreds or even thousands of dollars through autonomous agents like OpenClaw, making it unsustainable for Anthropic's finances and inference computing infrastructure.

Technically, Anthropic's own tools, such as Claude Code and Claude Cowork, are designed to improve the efficiency of their 'prompt cache,' which reuses previously processed text, whereas external tools like OpenClaw were sometimes not optimized for this mechanism. VentureBeat explains that these inefficient agents were reprocessing large amounts of data, threatening the stability of the service for a broader user base.

This new Agent SDK credit system links the costs of such inefficient program usage to dedicated credits rather than the entire subscription fee. In other words, if an inefficient agent consumes a large amount of tokens, it simply uses up the user's dedicated $20 to $200 allowance faster, and Anthropic will no longer have to continue absorbing the burden that exceeds the scope of the subscription plan.



On the other hand, XDA-Developers , which runs a developer forum, views this change critically, seeing it not as an 'addition of free credits,' but rather as a measure that diminishes the value of the game.

According to XDA-Developers, Agent SDK and claude -p were previously consumed from the subscription's usage limit, allowing users to use more tokens than they actually paid under the subsidized pricing structure. However, under the new system, once the dedicated credits are used up, any additional usage will be charged at the standard API rate.



XDA-Developers also points out that in Anthropic's efforts to separate coding use from AI interaction, it previously blocked third-party tools like OpenClaw from using subscription rate limits and instead required users to switch to API billing. One user, who hadn't used any third-party tools, was flagged by the detection logic because their git commit message contained the string 'HERMES.md,' and was charged $200.98 (approximately 31,700 yen) for API usage. XDA-Developers explains that Anthropic initially had no intention of refunding this charge, but did so after the case became widely known.

XDA-Developers sees this new dedicated credit system as a way to properly charge for third-party usage while avoiding the problem of sudden, exorbitant bills, but at the same time criticizes it, saying that developers are seeing it as 'trying to make a step backward look like good news.'

Related Posts:

in AI,   Web Service, Posted by log1i_yk