Llm Cost Advisor

LLM Cost Advisor (WHAT)

Recommend the right provider for the task at hand — balancing quality, cost, and privacy.

When to use

• User asks "which model should I use for X?"
• Before scheduling expensive loops (pr-babysitter, ci-sweeper)
• When dev-companion jobs are unexpectedly expensive
• When choosing between local and cloud providers for sensitive work

Out of scope

• Does NOT configure providers — that's done via env vars and loadenv
• Does NOT run LLM calls — it advises, the runner executes

Workflow

Step 1 — Check available providers:

devcompanion llm-status

Step 2 — Classify the task:

| Task type | Recommended tier | Why | |-----------|-----------------|-----| | Interactive code review, planning, debugging | Claude Sonnet / GPT-4o | Reasoning quality matters | | Batch labeling, summarization, triage | Claude Haiku / GPT-4o-mini | High volume, lower stakes | | Private/sensitive codebase work | Ollama or OpenCode | Stays local, zero cost | | Loop L1 (observation only) | Haiku / GPT-4o-mini | Low stakes, runs daily | | Loop L2 (PR creation) | Sonnet | Creates external artifacts | | Loop L3 (unattended) | Sonnet + verifier | Autonomous — quality critical |

Step 3 — Show cost estimate:

| Provider | Model | Est. per run | Monthly (1/day) | Monthly (96/day) | |----------|-------|-------------|-----------------|-----------------| | Anthropic | claude-3-5-sonnet | ~$0.05 | ~$1.50 | ~$144 | | Anthropic | claude-3-haiku | ~$0.005 | ~$0.15 | ~$14.40 | | OpenAI | gpt-4o | ~$0.04 | ~$1.20 | ~$115 | | OpenAI | gpt-4o-mini | ~$0.003 | ~$0.09 | ~$8.64 | | Ollama | (local) | $0 | $0 | $0 | | OpenCode | (local) | $0 | $0 | $0 |

Estimates based on ~2K input + ~1K output tokens per typical run.

Step 4 — Recommend with reasoning:

Give a clear recommendation: "For daily CNA issue triage (L1, observation only), use claude-3-haiku — ~$0.15/month vs $1.50/month for Sonnet, with no quality difference for labeling and summarization tasks."

Budget gate reminder

If the user is setting up loops, remind them to add limits.max_cost_usd to their job definitions to enable the runner's hard budget enforcement:

"limits": {
  "timeout_sec": 300,
  "max_cost_usd": 0.50
}

Anti-patterns

• Do not recommend cloud providers for repos marked confidential — suggest Ollama/OpenCode
• Do not recommend Haiku for L3 unattended loops — quality failures cascade
• Do not skip the llm-status check — recommend only available providers

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ulises-jeremias
• Source: ulises-jeremias/agent-toolkit
• License: MIT
• Homepage: https://ulises-jeremias.github.io/agent-toolkit-archive/

Install and usage instructions live in the source repository linked above.