IF YOU WANT TO GET AN AI ASSISTANT TO WORK FOR YOURSELF, IT'LL TAKE YOU LESS THAN THREE DAYS TO LOOK AT THE API BILL AND GET YOUR HAIR NUMB -- 100 MORE KNIVES A DAY, AND IT'LL COST YOU THOUSANDS A MONTH. I DON'T KNOW. THERE'S A LOT OF FRESHMAN WHITE UNDER THE LAST BIG POST OpenClawbut because token won't use it, you have to watch the ultra-priced ration
It was thought that open-source projects would be free for free, and it became clear that the software itself was free of charge and that the real gold swallower was the API call for the AI model. Every time you say a word, every time you run, every time you change a tool, you burn Token, and Token is real and silver。
After six months in the pit, I tried all the ways on x and github, and I finally took the monthly bill from the initial 300 to the level of stability, within 10, the extreme operation could even press five! All the whites that were shared with me today were hand-held dry products, at least 801 TP3T, and the whole thing could be done without knowing the complex code。
FIRST OF ALL, 901 TP3T'S MONEY CAME FROM BOTH THINGS
- Every call brings a lot of useless stuff
- Use the most expensive model to do the most basic work
NUMBER ONE: CHANGE THE MODEL! ONE MINUTE IS DONE
It's the most cost-effective, quickest move, none of it
The official default for Claude Opus is huge, and the input and output of Token is 15/25 knives per million, but to be honest, we don't use its power to talk, to alert, to change documents, to translate. Claude Sonnet is fully enough, with one third of the price cut directly; and one fifth of the cheaper Haiku, with Opus。
Change the default model to the Sonnet main force + Haiku to the bottom, not to be afraid of broken clothes, which is fully used daily:
bash
run
i'm sorry, openclaw config set `agents.defaults.model' -json'
“primary”: “antropic/claude-sonnet-4-5”,
“fallbacks”: [“anthropic/claude-haiku-4-5”]
}'
# RESET EFFECTIVE
openclawgaterestart
Progress is extremely cost-saving: the budget is more tightly and directly used for the national production Minimax M2.5, and the input is only 0.3 knife/million Token, which is 1/10 of the Sonnet, and the daily tasks are fully adequate. OpenClaw was originally supported, and even API Key was authorized without manual filling。
Strike two: QMD! Drop a nuclear weapon, direct 90% Token
It's White's work. It's done in two minutes
Most of the reason we spent money on it was that OpenClaw sent the entire memory file, all the chat records to AI, 90% for nothing。
QMD is dedicated to this: it helps you find what's most relevant to the problem at the local level, just a little bit to AI. Officially measured Token consumption is directly reduced by 90%-99%, which is 5-50 times faster, and AI's answer is more accurate and free from irrelevant information。
[Opplication] OpenClaw 2026.2.2 is directly built in and enabled, with no additional configuration, no cost
Number three: Cut off the ineffectual "noise" and delete all fixed money
The fixed cost of each call is the "bottom noise" of the system hint. The default file contains a whole bunch of things you can't use, burning up 10,000 Tokens every time, streamlining the basic consumption of a single call from 13,000 + Token to 300,000-5,000 Token。
It's very simple. Just do it
- Streamline AGENTS.md: Delete the rules of group chat without group chat, delete the TTS configuration without voice, leave only the functionality you can use, and reduce the target to 800 Token
- Streamline SOUL.md: Don't write a big character description. Just two or three sentences to explain what you want it to do
- Regular memory check: MEMORY.md directly archive the expired contents, not all piles
- Don't use one session at the end: regular new sessions, chat records are stacked and dozens of pieces burn once called. The chat box lost/status can see the current session. Token consumes
Number four: Take care of the walk-in. Don't let it burn
OpenClaw's heart beat function. A lot of people have a five-minute check on the mailbox and burn it all the time
Two steps
1. Change heart rate from the default of 10 minutes to 30 minutes or even 1 hour and version check to 1 day 1 time
bash run
it's an openclaw config set `agents.defaults.heart beat. every'30m'
SEVERAL SCHEDULED MISSIONS WERE CREATED TO COMBINE MAIL CHECK, CALENDAR READING AND PROCESSING INTO A “DAILY MORNING BULLETIN”, ONE RUN IN THE MORNING, ONE CALL WAS MADE AND THE COST OF 751 TP3T WAS DIRECTLY SAVED
Number five: work in separate places, do the right thing with the right model
Don't let an AI do anything, build two Agents, perfect for the more expensive and stupid:
- Main power split: Only long writing, knock code, complicated analysis with Sonet/Opus
- Lightweight split: Haiku/MiniMax, with only weather check, alerting, simple translation, with only the most streamlined configuration, and token with extremely low consumption
The sixth move is to save the wool
- Low usage: Google Gemini free floor, Flash model size enough to access OpenClaw, with little cost
- Stable: API bill exceeding 20 Knife/ Month, direct to Claude Pro subscription, much more than a single API payment
- With high-quality computers: Run the Ollama local model, API costs zero and is fully used on a daily basis
I'll pay you back
30 TIMES A DAY WITH BASE TIMED TASKS:
- Unoptimized use: 2 million Token per day, monthly bill 300-600 Knife
- Optimizing above: only 150,000-300,000 Tokens per day, monthly bills 10-25 knives, maximum operational pressure within 5 knives
901 TP3T. IT'S ALL PERSONAL
Finally, OpenClaw is really the best personal AI assistant at the moment, but free open source is not a zero cost. If you don't do it, you're a gold-eating beast。
You don't have to finish everything up there, just three moves and save half of your money