Anthropic has temporarily postponed the transition to a new pricing model for the Claude Agent SDK, which would have shifted from fixed subscriptions to pay-per-token usage via API. This decision was made amid concerns regarding a sharp and uncontrolled increase in costs for developers and users of tools utilizing the SDK.
What Happened
Anthropic has decided to maintain the current subscription model for using the Claude Agent SDK, canceling the planned transition to API-based pricing. Currently, users can continue working with the SDK under their existing Claude plans without additional costs. The decision aims to prevent a sudden increase in expenses for active users and developers, particularly clients of tools such as the Zed code editor.
Context
Existing Claude subscription models were not optimized for the specific workload patterns characteristic of autonomous AI agents, which generate massive volumes of requests. Transitioning to token-based billing could have created economic instability for those building agent-based workflows, prompting the company to reconsider the timing and approach to implementing the new system.
Why It Matters for the Industry
This situation serves as a signal of the inevitable transformation of the AI agent economy. The industry is moving toward a model where pay-per-token or specific agentic action billing becomes dominant; however, the current "economic immaturity" of these models requires the search for more balanced hybrid solutions that combine fixed subscriptions with variable computation costs.
Why It Matters for Users
For developers and users, this means maintaining cost predictability in the short term. There is a temporary window for prototyping and testing agentic functions without the risk of receiving an unexpected API bill while Anthropic develops a more sustainable pricing system.
What Is Not Yet Known / Limitations
There is uncertainty regarding exactly what type of pricing will be implemented in the future—whether it will be pure pay-per-token, a hybrid model, or specific plans tailored for agentic workloads.
Sources
Author
Look at AI, Editorial Team