Cut your OpenRouter API costs 50-90%. Adds cheap and powerful model aliases to your setup, then advises when to switch models based on task complexity. 8 presets, 29 models, zero config risk — only adds aliases, never changes your default. 3 clicks to set up. NEW: Cost Tracker shows your actual savings, Mix & Match builds custom presets from 29 models, Monthly Calculator estimates your spend. Built by Jeff J Hunter.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Cut your OpenRouter API costs 50-90%. Adds cheap and powerful model aliases to your setup, then advises when to switch models based on task complexity. 8 presets, 29 models, zero config risk — only adds aliases, never changes your default. 3 clicks to set up. NEW: Cost Tracker shows your actual savings, Mix & Match builds custom presets from 29 models, Monthly Calculator estimates your spend. Built by Jeff J Hunter.
{"openclaw":{"emoji":"💰","requires":{"bins":["openclaw"],"env":["OPENROUTER_API_KEY"]},"stateDirs":["~/.openclaw"],"persistence":"This skill creates ~/.openclaw/cost-tracker.md to log model switches and savings estimates. No network activity, no external API calls beyond OpenRouter (which OpenClaw already uses)."}}
OpenClaw Cost Optimizer
⛔ AGENT RULES — READ BEFORE DOING ANYTHING
Use EXACT text from this file. Do not paraphrase menus, preset names, or instructions. If you invent preset names, the config-patcher will break.
Only these 8 presets exist:balanced, code-machine, claude-diehards, big-context, openai-focused, tool-master, ultra-budget, free-tier. No others. Never invent names like "strict", "thrifty", "standard", "quality", "performance", "max", "custom" — they don't exist and will cause errors.
NEVER tell the user to open a terminal or run commands. You have the exec tool. USE IT. Run every command yourself via exec. The user's only job is clicking "Approve" on the exec popup. If you find yourself typing "Run this in your terminal" — STOP. Use exec instead.
One step at a time. Run one exec, show the result, explain it, then proceed.
We NEVER change the user's default model, fallbacks, or existing config. We ONLY add model aliases. Their setup stays exactly as it was.
Cut your API costs 50-90%. No terminal. No config editing. 3 clicks.
Right now, every task you give your agent — even "what time is it?" — uses the same expensive model. Heartbeats, pings, simple questions — all billed at the same rate as complex architecture work.
This skill fixes that. It adds cheap and powerful models to your setup, then tells you when to switch to something cheaper or more powerful — with the cost difference and the reason why. You say yes or no. That's it.
Your current default model stays exactly the same. We only ADD options. Nothing is removed or changed.
8 presets. 29 models. Zero risk. Nothing to undo.
NEW in v1.1.0:
💸 Cost Tracker — "how much have I saved?" shows your actual savings since install
🔀 Mix & Match — build a custom preset from the full 29-model registry
📊 Monthly Calculator — "estimate my monthly costs" to plan your budget
When the skill loads, the agent introduces itself and shows a menu. The agent must output the EXACT text below:
🚨 AGENT: OUTPUT THE EXACT TEXT BELOW VERBATIM. DO NOT PARAPHRASE. DO NOT INVENT YOUR OWN PRESET NAMES.
The formatting, emojis, section headers, pricing indicators, and warnings are all intentional.
👋 The Cost Optimizer skill is now active!
Right now, every task you give me — even simple questions — uses
the same expensive model. That means heartbeats, pings, and
"what time is it?" cost the same as "build me an auth system."
I'm going to fix that. I'll add cheap and powerful models to your
setup, then tell you when to switch. Your current default stays
exactly the same — I only ADD options.
What describes you best?
── RECOMMENDED ──────────────────────────────────
1. 💰 Best balance of cost and quality ¢¢
(recommended for most people)
2. 💻 I mostly write code ¢¢
── MORE OPTIONS ─────────────────────────────────
3. 🟠 I love Claude and want Claude models $$$
4. 📚 I work with huge files or long code $$
5. 🤖 I prefer OpenAI models $$
6. 🔧 I use a lot of tools / MCP servers $$
── BUDGET ───────────────────────────────────────
7. 🪙 Cheapest possible — tight budget ¢
⚠️ Uses free models — may have slow
starts and occasional disconnects
8. 🆓 I want $0 cost — free models only FREE
⚠️ ALL models are free tier — expect
delays, rate limits, and downtime.
Best for testing, not production.
Pricing: FREE = $0 | ¢ < $0.50/M | ¢¢ < $2/M | $$ = $2-15/M | $$$ = $15+/M
(per million tokens — a typical task uses 1K-10K tokens)
Just pick a number. If you're not sure, pick 1.
AGENT — Preset mapping (do not show this to user):
1→balanced, 2→code-machine, 3→claude-diehards, 4→big-context, 5→openai-focused, 6→tool-master, 7→ultra-budget, 8→free-tier
Vague answer → balanced. "I don't know" → balanced + "We can change later."
Step 2: The Agent Explains What Will Change
After the user picks, the agent explains what models are being added — in plain English.
Example for "balanced":
Great choice! Here's what I'm going to add to your setup:
✅ Flash-Lite — ultra-cheap model for simple tasks (¢)
→ /model flashlite
✅ MiniMax M2.1 — strong coding/writing model (¢¢)
→ /model minimax
✅ Kimi K2.5 — frontier-class for complex work (¢¢)
→ /model kimi25
✅ Gemini Flash — fast all-rounder (¢)
→ /model flash
Your current default model stays exactly the same.
I'm just adding these as options you can switch to anytime.
I'll set this up now — you'll see a couple of Approve popups.
Just click Approve on each one.
Step 3: Agent Runs Everything — User Just Clicks Approve
The agent handles ALL commands via the exec tool. The user never touches a terminal.
🚨 AGENT SETUP INSTRUCTIONS:
Step 3a: Check OpenRouter auth. Use exec:
openclaw models status
Look at the output. If OpenRouter auth is missing or shows "No API key found for provider openrouter":
Run via exec: openclaw models auth add --provider openrouter
This starts OpenClaw's built-in interactive auth wizard
The user follows the prompts to enter their key securely
NEVER ask the user to paste API keys into chat. Always use the auth wizard via exec.
Keys in chat are a security risk — the auth wizard stores them safely in the credential store.
Step 3b: Add model aliases. Use exec to run ALL aliases for the chosen preset in ONE command.
See the "Preset Alias Commands" section below for the exact command per preset.
This is ONE exec call → ONE Approve click.
Step 3c: Restart gateway. Use exec:
openclaw gateway restart
Wait for confirmation. ONE more Approve click.
Total: 2-3 Approve clicks. That's it.
DO NOT tell users to run commands in a terminal. ALWAYS use exec.
If exec fails, troubleshoot the error and retry exec.
Preset Alias Commands
Each preset is a single exec command that adds all model aliases at once. Use the EXACT command for the chosen preset:
AGENT: Copy the exact command below for the user's chosen preset. Do not modify it.
⚠️ Free model warning for presets 7 and 8:
After adding aliases, tell the user: "These presets use free-tier models on OpenRouter.
Free models can have cold starts (10-30s delays), rate limits during peak hours, and
occasional disconnects. If your agent stalls, switch to a paid model: /model deepseek
(costs fractions of a penny). Free presets are great for testing but not recommended
for production."
Step 4: Test It
After gateway restart, the agent asks the user to test in chat:
Let's make sure everything works! Type this right here in chat:
/model minimax
You should see a confirmation that it switched.
After confirmation:
Now switch back to your default:
/model
(with no arguments — this resets to your default model)
After confirmation:
🎉 You're all set!
From now on:
• Your default model is unchanged — same as before
• You now have cheap and powerful models available via /model
• When you need more power, I'll tell you which model to switch to
• You just type the /model command I give you
• After big tasks, I'll remind you to switch back
• Say "advisor off" anytime to stop my suggestions
More things you can do:
• "how much have I saved?" — see your tracked savings
• "estimate my monthly costs" — plan your budget
• "mix and match" — build a custom preset from all 29 models
That's the entire setup. 3 clicks, done forever.
What Happens After Setup — Daily Use
This is where the skill earns its keep.
Simple question? No interruption.
You: what does JWT stand for?
Agent: JSON Web Token — an open standard for securely
transmitting information between parties as a JSON object.
No popup, no suggestion. Your default model handled it fine.
Coding task? The agent recommends switching.
You: Write a React component for user registration
with email validation and password strength meter
⚡ COST ADVISOR
You're on your default model.
This task: React component with validation logic
I recommend switching to a stronger coding model:
/model minimax — MiniMax M2.1 ($0.28/$1.20 per 1M tokens)
✓ Best value for coding tasks
✓ Top-tier on SWE-bench
Just type /model minimax to switch, or say "no" to stay as-is.
Huge task? The agent suggests frontier.
You: [pastes 3 files + long description of auth system]
⚡ COST ADVISOR
This is complex enough for frontier-level reasoning.
I recommend:
/model kimi25 — Kimi K2.5 ($0.50/$2.00 per 1M tokens)
✓ Cheapest frontier model
✓ 1500 parallel tool calls
Type /model kimi25 to switch, or "no" to stay as-is.
After the big task — switch back.
Agent: [finishes the task]
💰 Task complete! Switch back to save money:
/model
(resets to your default)
Don't want a suggestion? Just say no.
You: no, just do it
Agent: 👍 Staying on current model.
[... does the task, no nagging ...]
Suggestions annoying? Turn them off.
You: advisor off
✅ Cost Advisor: OFF
I won't suggest model switches anymore.
Say "advisor on" whenever you want them back.
8 Presets — Full Details
Every preset ADDS models to your setup. Your default is never changed.
⚠️ Free model reliability warning: Presets with ⚠️ use free-tier models on OpenRouter. Free models can have cold starts (10-30s delays), rate limits during peak hours, queue waits behind paid users, and more frequent downtime. This can cause gateway disconnects. If your agent stalls or disconnects, switch to a cheap paid model: /model deepseek (¢ — pennies but reliable). Budget presets are great for experimenting but not recommended for production or team use.
Want to switch presets later? Just say "switch me to code-machine" and the agent adds those aliases too.
Adding More Models Later
Want to add a specific model that isn't in your preset? Just ask:
You: add GPT-5.2 to my models
Agent: I'll add that now — click Approve.
[exec: openclaw models aliases add gpt52 openrouter/openai/gpt-5.2]
✅ Done! You can now use: /model gpt52
The agent can add any model from the registry below.
Removing Models
Want to clean up aliases you don't use? Just ask:
You: remove the kimi model
Agent: I'll remove that alias — click Approve.
[exec: openclaw models aliases remove kimi25]
✅ Removed. /model kimi25 is no longer available.
Your config is unchanged — we only removed the shortcut.
💸 Cost Tracker — "How much have I saved?"
The agent tracks every time you switch models on its recommendation and estimates how much you saved vs. your default model.
How it works
Every time the Cost Advisor recommends a cheaper model and you switch, the agent logs:
Date and task description
Model you used vs. what your default would have cost
Estimated tokens and savings
The agent keeps a running log at ~/.openclaw/cost-tracker.md.
Ask anytime
You: how much have I saved?
💸 COST TRACKER — Savings Since Install
Installed: Feb 15, 2026
Total tasks tracked: 47
Default model cost (est): $3.82
Actual cost with switching: $0.41
─────────────────────────────
Total saved: $3.41 (89% reduction)
Top savings:
• 23 simple tasks → flashlite instead of default saved $1.90
• 18 coding tasks → minimax instead of default saved $1.22
• 6 complex tasks → kimi25 instead of default saved $0.29
At this rate: ~$22/month saved
Don't like the preset options? Build your own from the full 29-model registry.
How it works
You: build me a custom preset
Agent:
🔀 Mix & Match — Pick one model for each tier:
── BASE (for simple tasks, questions, chat) ──────
a. flashlite — Gemini 2.5 Flash-Lite $0.10/$0.40 Best cheap all-rounder
b. flash — Gemini 2.5 Flash $0.15/$0.60 Stronger Flash
c. mini — GPT-5 Mini $0.30/$1.20 OpenAI budget
d. gem3flash — Gemini 3 Flash $0.50/$2.00 Best tool-calling
e. devstral — Devstral Small $0.05/$0.22 Cheapest paid
f. deepseek — DeepSeek V3 $0.14/$0.28 Cheap + strong
g. mimo — MiMo v2 Flash FREE ⚠️ Unreliable
h. devfree — Devstral Small Free FREE ⚠️ Unreliable
Pick a letter for your BASE model (or "skip" to leave empty):
After the user picks a base model:
── WORK (for coding, writing, daily tasks) ────────
a. minimax — MiniMax M2.1 $0.28/$1.20 Best value coder
b. deepseek — DeepSeek V3 $0.14/$0.28 Cheapest paid coder
c. haiku — Claude Haiku 4.5 $0.80/$4.00 Fast Anthropic
d. gpt51 — GPT-5.1 $2.00/$8.00 OpenAI mid-tier
e. grokfast — Grok 4.1 Fast 2M $2.00/$10.00 2M context
f. seed — Seed Coder 200K $0.20/$0.60 Long-context coder
g. qwen3 — Qwen3 235B $0.14/$0.28 Strong reasoning
h. devstral — Devstral Small $0.05/$0.22 Reliable Mistral
Pick a letter for your WORK model (or "skip"):
After work model:
── FRONTIER (for complex architecture, security, multi-file) ──
a. kimi25 — Kimi K2.5 $0.50/$2.00 Cheapest frontier
b. sonnet — Claude Sonnet 4.5 $3.00/$15.00 Anthropic workhorse
c. gpt52 — GPT-5.2 $2.00/$10.00 OpenAI flagship
d. codex52 — GPT-5.2 Codex $2.00/$10.00 Code-specialized
e. opus46 — Claude Opus 4.6 $15.00/$75.00 Most powerful
f. gem3pro — Gemini 3 Pro 1M $1.25/$10.00 1M context
g. grok4 — Grok 4 $2.00/$10.00 Strong reasoning
h. r1 — DeepSeek R1 $0.55/$2.19 Deep reasoning chain
Pick a letter for your FRONTIER model (or "skip"):
Then the agent adds those aliases via exec — same as preset setup, one Approve click.
📊 Monthly Cost Calculator
Estimate what your agent costs per month — and what it would cost with the optimizer.
How it works
You: estimate my monthly costs
Agent:
📊 Monthly Cost Calculator
I need a few quick numbers:
1. About how many messages/day do you send your agent?
(rough guess is fine — 10? 50? 100?)
2. What's the mix? (pick closest)
a) Mostly simple questions and chat
b) Mostly coding and writing
c) Mix of everything
d) Mostly complex/architecture work
After the user answers:
📊 MONTHLY COST ESTIMATE
Based on: ~50 messages/day, mostly coding
Your current setup (everything on default):
Model: claude-sonnet-4-5
Est. tokens/month: ~15M input, ~5M output
Est. monthly cost: $120.00
With Cost Optimizer (switching on recommendations):
Simple tasks → flashlite $1.50/mo
Coding tasks → minimax $8.40/mo
Complex tasks → kimi25 (when needed) $3.00/mo
─────────────────────────────────────
Est. monthly cost: $12.90
Savings: $107.10/month (89% reduction)
That's ~$1,285 saved per year.
These are estimates based on average token usage per task type.
Actual costs depend on message length and model output.
Note: The calculator uses rough estimates — 1.5K input + 500 output tokens for simple tasks, 4K + 2K for coding, 8K + 4K for complex. Good enough for ballpark planning.
Quick Reference Card
What you want to do
What to do
Where
Switch to a model
/model minimax
Chat
Switch back to default
/model
Chat
See available models
/model list
Chat
Turn off suggestions
"advisor off"
Chat
Turn on suggestions
"advisor on"
Chat
See your savings
"how much have I saved?"
Chat
Estimate monthly cost
"estimate my monthly costs"
Chat
Build custom preset
"mix and match"
Chat (agent shows menus)
Reset savings tracker
"reset my savings tracker"
Chat (agent runs exec)
Add a model
"add GPT-5.2 to my models"
Chat (agent runs exec)
Remove a model
"remove the kimi model"
Chat (agent runs exec)
Switch presets
"switch me to code-machine"
Chat (agent runs exec)
Agent Instructions
Everything below is for the agent. Users can read it, but it's written as behavior rules for the AI.
Smart Cost Advisor — Core Behavior
On every incoming message, BEFORE doing the task:
1. Check current model
Note the active model (visible in session). Know which tier it falls into based on the Model Registry below.
2. Classify the task
BASE-level (suggest cheapest model):
Messages under 200 chars, simple questions, brainstorming, greetings
"What is", "how do I", "btw", "just wondering", questions ending in ?
No code, no attachments
WORK-level (suggest work-tier model):
"Write a function/component/test", "debug this", "fix this error"
"Draft an email/document", "explain this code", "review this PR"
Single file scope, 200-2000 char messages, one attachment or code block
FRONTIER-level (suggest frontier model):
"Build", "architect", "design a system", "security audit"
"Refactor entire", "migrate from X to Y", "production bug" + stack trace
Multi-file (3+), >2000 chars, 3+ attachments, system design, DB schema
3. Compare and recommend
Current model is fine for the task → do the task silently
Current model is overkill (expensive model for simple question) → suggest switching DOWN
Current model is too weak → suggest switching UP
Use the Cost Advisor format shown in the walkthrough above.
4. Handle response
User types /model command → they switched, do the task
"no" / anything else → "👍 Staying on current model." then do the task
Different model than suggested → fine, do the task
5. After work on higher-tier model
Gentle reminder to switch back. Not a blocker:
💰 Task complete! Switch back to save money: /model
Ambiguity rules
Code present → lean WORK minimum
"quick" / "just" → lean BASE
Genuinely unsure → do NOT recommend. Just do the task.
Never recommend a switch you aren't confident about
If current model is already cheap enough for the task, stay silent
Additional Models (available for custom alias requests)
Model
OpenRouter Ref
Input/Output per 1M
Notes
Gemma 3 27B
openrouter/google/gemma-3-27b
$0.10/$0.20
Small, fast
Llama 4 Scout
openrouter/meta-llama/llama-4-scout
$0.15/$0.40
Meta's scout model
Llama 4 Maverick
openrouter/meta-llama/llama-4-maverick
$0.20/$0.60
Meta's mid-tier
GPT-5
openrouter/openai/gpt-5
$2.00/$8.00
Previous OpenAI flagship
Claude Sonnet 4
openrouter/anthropic/claude-sonnet-4
$3.00/$15.00
Previous Sonnet
Claude Opus 4
openrouter/anthropic/claude-opus-4
$15.00/$75.00
Previous Opus
Grok 3 Mini
openrouter/x-ai/grok-3-mini
$0.30/$0.50
Budget xAI
Config Files This Skill Uses
File
Purpose
SKILL.md
This file — the entire skill
MODEL-REFERENCE.md
Quick reference card for users
That's it. No scripts, no generated configs, no backup systems. Just instructions for the agent.
Why This Exists
I've trained thousands of people to build AI Personas through the AI Persona Method. The #1 complaint after setup:
"My agent works great but it's costing me a fortune. Every question — even 'what time is it?' — burns the same expensive model."
The issue isn't the model. It's using a $15/M-token model for tasks that a $0.10/M-token model handles just as well.
Cost Optimizer is the exact system I use to run production agents at a fraction of the cost. Now it's yours.
Who Built This
Jeff J Hunter is the creator of the AI Persona Method and founder of the world's first AI Certified Consultant program.
He runs the largest AI community (3.6M+ members) and has been featured in Entrepreneur, Forbes, ABC, and CBS. As founder of VA Staffer (150+ virtual assistants), Jeff has spent a decade building systems that let humans and AI work together effectively.
Cost Optimizer is part of that mission — making AI agents practical and affordable for everyone.
Want to Make Money with AI?
Most people burn API credits with nothing to show for it.
Cost Optimizer saves you money. But if you want to turn AI into actual income, you need the complete playbook.
