omp ships 32 tools, and every enabled tool's definition rides along in context — the quiet fixed cost per request. Trim the tool surface to what you use, exploit its intent slots (smol for routine steps), and lean on hashline edits, which exist precisely to cut edit-token waste.
If you never use the browser or DAP tools, disabling them removes their definitions from every request. On a 32-tool surface this is a real per-call saving.
omp's intent routing lets routine steps run on a cheap model automatically. Set smol to a value model and let default escalate only when needed.
Hash-anchored edits mean the model references content instead of retyping it — fewer output tokens and fewer corrupted-edit retries. Prefer them over full-file rewrites in your workflow.
The provider-agnostic tactics (prompt caching, retry budgets, batch APIs) are in the general playbook.
The tool surface: every enabled tool's definition is part of the request context. Disabling unused tools and caching the stable preamble cuts the fixed overhead per call.
Yes, if configured: routing routine steps to a cheap model in the smol slot while keeping default/plan on stronger models is exactly the route-by-task-size tactic, built into the agent.
The structural version of all of this: run Oh My Pi on a flat monthly price with unlimited tokens, and the bill stops being a variable to manage. 2-minute omp setup → · Best models for omp →