Anthropic has paused a planned update that would have changed how developers are charged for its Claude Agent SDK, keeping the current billing structure in place while it rethinks how pricing should work for AI agents.
The update, first reported by Ars Technica, was part of a broader plan to introduce token-based billing for more advanced usage inside Anthropic’s developer ecosystem. That approach would have tied costs directly to how much text the model processes, including both user prompts and AI-generated outputs.
To understand what was at stake, token billing is already the standard pricing model across most AI systems today. Every interaction with a model is broken into tokens, which are small units of text. The more tokens a request uses, the more it costs. It is a system designed to match pricing with compute usage, especially for APIs where workloads can vary widely.
Anthropic’s earlier plan would have extended that model more aggressively into AI agents , tools built on top of models that do not just respond to prompts but also execute tasks, run workflows, and operate continuously in the background. These systems tend to consume far more tokens than standard chat interactions because they carry out multiple steps without stopping.
That is where the challenge began to show.
Unlike simple queries, agent-based systems can run long sequences of actions, trigger repeated model calls, and stay active for extended periods. Developers building on top of these tools have increasingly raised concerns that token-based billing at full scale could make costs unpredictable, especially for production systems.
Because of that, Anthropic has now decided to pause the planned billing change rather than roll it out as originally expected. The company says it is still reviewing how best to structure pricing for developer tools that rely heavily on autonomous agent activity.
For now, users of the Claude Agent SDK will continue under the existing billing model, with no immediate change to pricing or access.
This decision sits inside a wider tension in the AI industry. As models move from simple chat interfaces into systems that perform work on their own, pricing becomes harder to define. What used to be predictable text generation is now turning into continuous, multi-step execution that behaves more like software infrastructure than a chatbot.
Token billing still works well for traditional API usage, but AI agents are forcing companies to reconsider whether that model alone is enough to cover how these systems are actually used in real environments.
Anthropic has not given a timeline for when a revised pricing approach might return. For now, the company is holding the line while the industry continues to adjust to what AI agents are becoming , not just tools that respond, but systems that act.
Read also: Binance risks EU exit as Greece reviews MiCA license