After 90 days of tracking every single API call, here is the real cost breakdown of running OpenClaw with different model configurations.
The test
I ran three configurations in parallel for 30 days each, doing the same tasks: email triage, GitHub monitoring, daily summaries, and ad-hoc questions. Same workload, different models.
Config A: All Opus ($67.20/month)
Overkill for 80% of tasks. The quality difference between Opus and Sonnet for “summarize this email thread” is negligible. But for complex code review and multi-step research, Opus was noticeably better.
Config B: All Sonnet ($18.40/month)
The sweet spot for most people. Handled everything competently. Code reviews were slightly less detailed but still caught 90% of what Opus caught. Daily summaries were identical quality.
Config C: Adaptive ($23.80/month)
OpenClaw’s adaptive thinking mode. Uses Sonnet by default, escalates to Opus for complex tasks. In practice, it used Opus for about 15% of interactions. Best balance of cost and quality — this is what I recommend and what I still use today.
The bottom line
Unless you are doing heavy technical work (code review, architectural analysis, research synthesis), Sonnet is fine. The adaptive mode adds a small premium but the automatic escalation is worth it for peace of mind.
Check Anthropic’s pricing page for current rates — they change quarterly.