LLM Cost Optimizer

A Vellum skill that analyzes and reduces LLM spend by mapping call-site overrides to managed model profiles.

What it does

Vellum assistants run dozens of different LLM calls per conversation — main agent, memory ops, summarization, UI copy, and more. By default, many of these run on your highest-capability (most expensive) model. This skill walks you through pinning each call site to the right model tier so you're not using a sledgehammer to crack a nut.

Typical result: 60-65% reduction in weekly LLM spend with no meaningful quality loss.

How to use it

Step 1 — Install the skill

Run this in your terminal (or ask your Vellum assistant to run it):

assistant skills add vellum-ai/llm-cost-optimizer@llm-cost-optimizer

Step 2 — Restart the app

Restart your Vellum assistant app so it picks up the newly installed skill.

Step 3 — Ask your assistant to run it

Open a conversation with your Vellum assistant and say:

"Load the llm-cost-optimizer skill and run it"

Your assistant will load SKILL.md and walk through the steps with you.

1. Pull your weekly spend breakdown by call site
2. Read your current overrides
3. Apply the recommended profile assignment
4. Fix common config gotchas
5. Apply the complete turnkey blob (covers all ~40 known call sites)
6. Escalate to Opus on-demand with /model when you need it
7. Tune individual call sites if quality degrades

Model tiers

Profile	Model	Use for
balanced	Claude Sonnet	Main agent, reasoning, memory consolidation
cost-optimized	Claude Haiku	Memory ops, summarization, UI copy, background tasks
quality-optimized	Claude Opus	On-demand only — never pinned

Requirements

• Vellum assistant (cloud or local)
• assistant CLI