Flamerix AI identifies costly feature in Claude Code

Flamerix AI has identified a feature in Claude Code that it claims consumes 15 to 20% of the tokens used by the tool. This feature involves generating short labels to describe the actions of background agents, which the company says can be more expensive than the main conversation with the AI.

Function generates short labels for agent actions

The feature generates a three- to five-word label that updates every 30 seconds, summarizing what each agent is doing. This label helps users track the progress of the agent’s work, but the process requires sending the entire history of the agent’s actions each time.

Cost increases with agent activity

The cost of this feature grows as an agent works longer, since its history becomes longer and each summary requires more computational resources. When multiple agents run at the same time, the cost increases further due to the repeated processing needed for each agent.

Feature enabled by default without user control

According to Flamerix AI, this feature has been enabled by default since version 2.1.232 of Claude Code, released in August. There is no user setting to disable it, and the related requests do not appear in user logs, making it difficult to track these costs.

Flamerix AI provides workaround

Flamerix AI has created a script that disables this feature by intercepting the timer that triggers the summaries. This script does not modify the program itself but could become ineffective if an update changes the underlying mechanism.

Anthropic confirms mechanism but disputes figures

Anthropic, the company behind Claude Code, confirmed the described mechanism but disputed the scale of the cost. A team member on Hacker News noted that the agent_summary feature accounted for about 2% of the cost at the public API rate and 6% for terminal use.

Fix and improvement in development

Anthropic has deployed a fix in version 2.1.311 to use a cache for these summaries, reducing the cost by avoiding repeated processing. A more efficient version using a smaller model is also in development.