There is a habit in this business of looking for the one best model and then pointing everything at it. Web Experts has spent a lot of hours experimenting with OpenClaw, and the conclusion is that the habit is expensive and it does not even produce better work.
The cost-effective setup is not about finding a winner. It is about assigning the right model to the right job and letting cheap models do the boring parts.
The Routing Setup
For default chat and general back-and-forth, GPT-4.0-nano handles it. Planning and deep reasoning go to DeepSeek-R1, which is where the real thinking should be happening anyway.
Code writing and refactoring split two ways. GPT-5.3 takes the high-stakes work, and DeepSeek-Coder v2 covers the cost-efficient side when the task does not justify the premium. Code review and sanity checks go to GPT-4.1, which is a different set of eyes on the same problem.
Shell and local execution run on GPT-OSS, kept local. Summarization and cleanup go back to GPT-4.0-nano, and DeepSeek-V3 sits underneath everything as the cheap general fallback.
Why It Keeps Costs Down
Three things are doing the work here. The setup maximizes cached inputs, it keeps execution local instead of paying a vendor to run a shell command, and it reserves the expensive models only for work that actually benefits from them.
That last point is the one most teams miss. Plenty of agent traffic is cleanup, restating, and shuffling text around, and paying top-tier prices for that is just a donation.
What This Buys You
The practical result is that an agent stack stops being something you nervously watch a billing dashboard for. You get to leave it running, which is the entire point of building agents in the first place.
Web Experts is not claiming this is the final answer for every shop. It is the setup that came out of a lot of trial and error, and the underlying principle holds even as the model names change: maximum output, minimum spend.