Recently, the AI engineering world has been talking about the new Claude Agent SDK usage rules presented by Anthropic.
Context: starting June 15th, any programmatic usage of Claude Code by Claude subscribers will run on a separate credit which is doled out monthly. This is separate from your normal subscription usage included in your plan.
On one hand, it is nice to receive clarity. Now third-party developers can actually support Claude Code subscribers who don't want to pay for API credits seperately. And it could be thought of as adding additional value (you're essentially getting an extra $100 on top of your normal usage limits). However, I still believe it misses the mark for a few reasons.
First, claude -p usage is now run on this credit. Previously, users could use this run-mode to
engage in some basic agent automations. It was quite handy and covered maybe 80% of automation use-cases. But these
capabilities are now severely limited. Once the credit is used, you're in "pay-as-you-go" territory for the
remainder of the month. In that sense, this change reduces subscription value, potentially significantly.
Next, we still must use claude -p or the Claude Agent SDK to make use of this credit. Oftentimes, this is not ideal.
With Meridian (via NullClaw) and Hermes (via its Claude Code provider), I've seen the underlying model misidentify itself
as Claude Code. This makes sense as commands are being run through the Claude Code agent.
Hermes uses ACP, Meridian uses Claude Agent SDK, but both create a small abstraction layer between the agent and the orchestration/substrate engine. In this abstraction, issues fester. If we could instead use an API key, these hiccups would be resolved.
Finally, and this is a smaller gripe, it essentially introduces a third bucket for credits. We have the subscription usage limits, the Claude SDK credit, and API credits (which are managed under a separate account). I would prefer it if the last two credits were at least combined...
Ultimately, the clarity is welcome, but it feels like this change introduces more new problems than it solves. It also feels like more evidence we're exiting the "free lunch" era of AI, which is in itself disappointing.