Developers Warn of High OpenClaw API Bills
Users of the open-source AI agent OpenClaw are reporting unexpectedly high API bills, with some developers seeing costs of several hundred dollars in just a few days. One report cited a monthly cost projection of $623 for running three agents on the Opus model. The discussions serve as a caution for developers to closely monitor API usage and implement cost controls when experimenting with autonomous agent frameworks.
- The primary drivers of high costs are technical defaults within the OpenClaw framework, not the complexity of the tasks being performed. Key issues include sending the entire conversation history with every new request, storing large tool outputs in logs, and repeatedly resending complex system prompts. One user reported that their session context alone grew to occupy over 56% of a 400,000-token window. - Anthropic's Claude 3 Opus, the model mentioned in one high-cost report, is priced at $15 per million input tokens and $75 per million output tokens. In one extreme case, a tech blogger accumulated a $3,600 bill from running the agent. More typical costs for active agents with frequent checks can range from $270 to $540 per month. - OpenClaw is an open-source framework that allows AI models to execute tasks on a user's machine, such as running shell commands, controlling a browser, and managing files, often triggered via messaging apps like WhatsApp or Slack. It is one of the fastest-growing open-source projects, having gained over 100,000 GitHub stars in under a week after its launch in late January 2026. - Cost-control strategies include routing simple tasks to cheaper, smaller models, preventing agents from running unattended, and using local models for routine "heartbeat" checks to see if there are tasks to perform. One developer built a tool called ClawWatcher specifically to add budget controls and automatic agent pausing when spending caps are hit. - Beyond cost, the framework's power introduces security risks; a high-severity vulnerability (CVE-2026-25253) was discovered that could allow a malicious website to gain remote code execution on the user's machine. Developers are advised to run the agent in an isolated environment and carefully audit any third-party skills before installation. - In response to cost and security concerns, a number of alternatives have emerged. These include Nanobot, a lightweight alternative with only 4,000 lines of code compared to OpenClaw's 430,000+; NanoClaw, which focuses on security by running the agent inside a sandboxed container; and Claude Code, which is specifically tailored for developer use cases.
