Anthropic splits Claude API limits, cutting developer throughput sharply
Original: Claude subscription changes coverage of `claude -p`
Source: x.com ↗
Who: Posted by ramoz on Hacker News, linking to an X trending-topic summary aggregating reactions to an Anthropic policy change. Theo (CEO of T3.gg, a developer-tooling company) is among the most prominent voices cited in the thread.
What's new: Anthropic is splitting Claude's subscription usage into two separate pools. The existing pool continues to cover normal back-and-forth chat. A new, separate pool — measured in fixed monthly credits rather than unlimited or soft-throttled use — now governs programmatic access, meaning calls made through code rather than a browser window. The credit amounts attached to each plan are $20 for Pro and up to $200 for higher-tier plans.
How it works: The split is enforced at the level of how the request arrives. Calls made through the — where software is driving Claude autonomously rather than a person typing — are now metered against the new credit budget. Interactive sessions are unaffected. Once a developer burns through their monthly credit allotment, they are cut off until the next billing cycle or must purchase more.
The numbers: Theo describes the effective cut as a 40x reduction in how much programmatic work a paid subscriber can do. Some developers report exhausting their entire monthly credit budget in under an hour of normal automated usage — suggesting the new caps are calibrated for light experimentation, not production workloads.
Why it matters: Developers who had built automated pipelines, internal tools, or early-stage products on top of Claude's API through a flat subscription now face a hard ceiling that did not exist when they made that engineering bet. Anthropic's stated rationale — protecting ordinary chat users from having their quotas crowded out by heavy programmatic traffic — is plausible as a product decision, but the abrupt change without a long migration window is what has generated most of the anger. The practical result is that some builders are migrating workloads to competing models, with OpenAI named explicitly in the thread as a destination.