OpenClaw has been popping up everywhere lately if you are paying attention to AI integration or agent automation. Web Experts spent part of last week running it in a sandbox to see whether it makes sense for real client environments.
The first reaction was honestly excitement. The second was concern, once we looked closely at the token burn.
What It Actually Does
OpenClaw is effectively Claude with hands. It can execute shell commands, manage local file systems, and automate browser workflows across platforms.
From a pure capability standpoint, it is one of the most powerful agentic frameworks we have tested so far. That power comes with a serious economic tradeoff, and the tradeoff is not obvious until you watch the meter.
The Loop Is The Cost
The architecture is deeply agent-native and relies on a continuous Plan, Observe, Act, Verify loop. To complete even a simple task, the agent does not make a single model call.
It continuously reinjects the entire session history and all tool definitions back into the context window at every step. You are not paying for what you just asked. You are paying to reread everything that came before it, over and over again.
That is where the burn gets aggressive. In longer sessions, the agent resends full skill definitions and massive conversation histories with every action.
How Fast It Adds Up
Without strict guardrails, an agent that misinterprets a task can chew through hundreds of dollars in API credits in minutes. There is no natural stopping point when the loop is convinced it is making progress.
And once the agent ingests large tool outputs, like directory trees or log files, that bloat becomes permanent for the life of the session. Every future interaction in that session costs more than the one before it.
Routing Is The Fix
Right now, model routing feels like the only viable way to make this sustainable. Web Experts has been experimenting with routing lightweight coordination and status checks to local or open source models while reserving high-reasoning models for the moments that actually require them.
That alone drops the operating cost floor dramatically. It is not elegant, but it turns an unpredictable bill into something you can plan around.
Why We Skip DeepSeek
There has also been a lot of momentum around DeepSeek because of its price-to-performance ratio. The reasoning quality is impressive, and we understand why people are drawn to it.
We avoid it for client work due to ongoing concerns around data privacy, factual reliability, and opaque filtering behavior. For most clients, data sovereignty matters more than token savings.
Structural Risk, Not A Line Item
The bottom line from our sandbox testing is that OpenClaw's autonomous capabilities are the most advanced we have seen this year, but out of the box efficiency is poor. In client-facing or enterprise environments, token burn is not a line item. It is a structural risk.
In 2026, success with agentic systems is not just about getting AI to act. It is about getting it to act predictably, responsibly, and without quietly lighting money on fire.