Descriptions:
Running every task — from 30-minute heartbeat pings to complex multi-file refactors — through a single top-tier model like Claude Opus is one of the costliest default behaviors in OpenClaw deployments. In this tutorial, VelvetShark walks through a single config change that implements model tiering: routing heartbeats to Gemini 2.5 Flash-Lite at $0.50 per million tokens and sub-agent work to DeepSeek R1 at $2.74 per million tokens, while keeping Opus ($30/M) reserved for tasks that genuinely need frontier reasoning.
The walkthrough covers the exact JSON structure in ~/.openclaw/openclaw.json, including heartbeat model assignment, sub-agent routing, and a cross-provider fallback chain that defaults to GPT-5.2 — not Sonnet — when Anthropic hits rate limits, since a provider-wide slowdown would affect all Anthropic models equally. Model aliases (opus, sonnet, flash, ds) enable quick in-session switching via /model commands. A companion savings calculator at calculator.vlvt.sh puts the numbers in context: a power user running 48 heartbeats per day drops from roughly $943 to $347 per month; heavy multi-agent setups fall from ~$3,000 to ~$1,000. OpenRouter’s automatic complexity-based routing is mentioned as a lower-configuration alternative for those who prefer not to manage tiers manually.
📺 Source: VelvetShark · Published February 04, 2026
🏷️ Format: Tutorial Demo
