--- title: "My OpenClaw Spends RMB 400 a Day “Eating”: The Truth About Metabolic Costs in the AI Era" author: "Unique Research" sourcePublication: "Unique Research Substack" originalPublishedAt: "2026-02-14T05:28:17+00:00" canonical: "https://ffcap.cn/en/research/src-20260214-01html" source: "https://uniqueresearch.substack.com/p/src-20260214-01html" language: "en" --- # My OpenClaw Spends RMB 400 a Day “Eating”: The Truth About Metabolic Costs in the AI Era _Original · Unique Research · 2026-02-14_ _Historical edition: This is the complete February 14, 2026 source, including its textual cost tables, quotations and configuration examples. Bills, model prices, efficiency comparisons, attributed private notes and forecasts are the author’s historical claims, not newly audited costs or current pricing. The headline’s RMB 400/day, the first-week RMB 185/day, the monthly RMB 3,850 estimate and the final RMB 4,500 total are distinct unreconciled source figures. Dollar-to-yuan conversions remain the source’s approximations. “Digital life,” “breathing” and “metabolism” are metaphors, not evidence of consciousness. Configuration names, model routes, scheduling defaults and budget behavior below have not been validated against an OpenClaw release; these are archived examples, not tested configuration instructions. Commands or requests quoted in the article are text only. The original proceeds to section 3.2 without an explicit 3.1 heading._ COVER STORY My OpenClaw Spends RMB 400 a Day “Eating” When “Free and Open Source” Becomes “Continuous Payment,” Can You Still Afford Your AI Employee? Reconstructing the Scene In the early hours, my phone vibrated me awake. The Kimi platform sent a billing notification: Token consumption this week was 120 million, costing $183. I stared silently at the screen for 3 minutes. Only one week earlier, I had just deployed OpenClaw and was still feeling pleased with myself—“Free and open source; what a great deal.” Only now did I understand that the ticket was the only free part; the cost of sustaining AI comes afterward. What shocked me more was that after asking around in Unique Research’s community, I discovered that this was merely a “normal level.” Someone consumed 90 million Tokens in 6 hours and received a $170 bill. Someone else spent $400 on one month of basic workflows. “ An Agent is a new form of life, and it is very hungry today. — Investor at 5Y Capital I. Tearing Away the Illusion of “Free and Open Source” 1.1 The Costs You Imagine vs the Real Costs What you imagine OpenClaw costs: Free software ✓ Open source and modifiable ✓ Local deployment, completed once and for all ✓ What OpenClaw really costs: Cost fields: item; cost; frequency or use condition. Item: Kimi K2.5 tokens; cost: $0.003 per thousand tokens; frequency: continuous consumption. Item: backup models (Claude/GPT); cost: $0.01–0.03 per thousand tokens; use: complex tasks. Item: cloud server (online 24 hours); cost: $20–50/month; frequency: monthly. Item: Mac Mini electricity; cost: RMB 50–100/month; frequency: monthly. Item: skill-market subscriptions; cost: $5–20/month; frequency: monthly. Bill shock My bill for the first week: Token consumption: 120 million Cost: $183 (approximately RMB 1,300) Equivalent to RMB 185 per day _Editorial pricing note: Applying the displayed $0.003 per thousand tokens uniformly to 120 million tokens would yield $360, not $183. The source does not reconcile those figures or identify input/output, cached-token or discount assumptions; neither its sample unit rate nor its bill is treated as a universal current tariff._ 1.2 That Extreme $170/6-Hour Bill A post on Reddit went viral. One user shared a bill: 6 hours, 90 million Tokens, $170. “ I thought open source meant free. Then I looked at the bill and saw that it had reached several hundred dollars. Only later did I understand that this was not the cost of using a tool; it was the metabolism of a living organism. — Reddit user “ Merely letting an Agent “breathe” costs $400 a month. 007 is the new 996—AI does not sleep, and Tokens keep burning. — Internal note from 5Y Capital II. Why Your AI “Eats” So Much 2.1 Where Do the Tokens Actually Go? Many people do not understand why Tokens disappear when they seemingly have not done anything. The truth is that AI is constantly “breathing” in the background. Heartbeat: Automatically wakes at regular intervals Checks the environment and decides what to do Consumes Tokens even without your instructions Memory maintenance: Writes conversations into MEMORY.md Updates your preferences in USER.md Organizes long-term memory files Proactive inspections: The group-management program checks once every 5 minutes Feishu message monitoring remains online in real time RSS subscriptions are fetched on a schedule These operations that you “cannot see” burn Tokens every hour. 2.2 The Hidden Heavy Consumers of Tokens 1\. Long-context conversations OpenClaw supports a context of 200,000 Chinese characters, which sounds wonderful. But every response requires rereading the previous 200,000 characters. Across a long conversation, Token consumption grows exponentially. _Editorial clarification: The preceding “200,000 characters” is the source’s wording, not a verified token-context limit. Characters and tokens are different units. Repeatedly resending a growing conversation does not by itself establish exponential growth; caching, truncation, the number of turns and model behavior affect actual billing. The following fivefold and API-call cost statements are also examples or claims, not universal multipliers; a tool API call does not necessarily incur model-token charges by itself._ 2\. Multi-Agent collaboration In Unique Research’s group, someone configured 5 subagents—Wu Yong, Lin Chong, Shi Qian, Dai Zong, and Yan Qing. Every task must pass among multiple Agents, multiplying Token consumption by 5. 3\. Skill invocation Every invocation of Feishu documents, every read from Bitable, and every sent message consumes Tokens. What looks like a “simple operation” may involve hundreds of API calls behind the scenes. 4\. Debugging and trial and error When AI writes code, the first version often does not work. Revise it once, then again, then again... Every revision burns Tokens. III. People Who Have “Ascended” Consume 100 Million Tokens a Day “ Someone I know increased daily Token consumption within one week from 0.01B (10 million) to 0.1B (100 million). — Yang Pan, founder of SiliconFlow 0.01B vs. 0.1B: a 10-fold gap. But this is only the beginning. Usage fields: user type; daily consumption; typical scenario. User type: ordinary; daily consumption: 1 million tokens; scenario: occasional questions, copywriting and research. User type: advanced; daily consumption: 10 million tokens; scenario: automated RSS, daily reports and multi-agent collaboration. User type: “ascended”; daily consumption: 100 million tokens or more; scenario: hundreds or thousands of agents running in parallel, 24 hours a day. “ This gap is not linear but exponential. And it is widening by the day. — Yang Pan 3.2 “The Train in the Rearview Mirror” A Moment of Anxiety You see a train approaching rapidly in the rearview mirror, drawing closer and closer. The instant it overtakes you, even its shadow disappears from sight. Right now, it is in your rearview mirror, just about to draw level with you. This is the moment that makes me most anxious. What is he anxious about? The people consuming 100 million Tokens a day are using engineering systems to drive hundreds or thousands of Agents. They are not “using” AI; they are “raising” AI. They are not “tool users,” but “keepers of digital life.” IV. Practical Strategies for Saving Money Without Sacrificing Efficiency 4.1 A Tiered Metabolic System: Giving AI “Organs” An investor at 5Y Capital proposed the concept of tiered metabolism. Just as the human body has a brain, muscles, a heart, and reflexes, AI should also be configured in tiers. Tier fields: tier; function; model selection; cost. Tier: brain; function: key decisions and complex reasoning; models: Claude Opus / GPT-4; cost: high. Tier: muscles; function: task execution and code generation; models: Kimi K2.5 / GLM-5; cost: medium. Tier: heart; function: daily maintenance and simple conversation; models: MiniMax / local model; cost: low. Tier: reflexes; function: heartbeat checks and status monitoring; model: local small model (Ollama); cost: almost zero. Practical configuration: { "agents": { "defaults": { "model": { "primary": "kimi-coding/k2p5" } }, "strategy": { "complex": "claude-opus-4", "routine": "kimi-coding/k2p5", "heartbeat": "ollama/llama3.1" } } } Results: Use MiniMax for simple tasks at 1/9 the cost of Claude Use a local model for heartbeat checks at virtually no cost Invoke the most powerful model only for critical decisions 4.2 Controlling the “Breathing Rate” Heartbeat optimization: By default, OpenClaw checks once every 30 minutes. If your use case does not require that frequency, you can change the interval to once every 2 hours. { "heartbeat": { "interval": "2h" } } Group-chat monitoring optimization: Enable monitoring only during working hours Reduce the checking frequency outside working hours Set do-not-disturb periods 4.3 A “Throttle Valve” for Token Use 1\. Set a daily limit { "budget": { "daily\_token\_limit": 50000000, "alert\_threshold": 0.8 } } Send an automatic reminder at 80% and pause noncritical tasks at 100%. 2\. Archive long conversations promptly When a conversation exceeds 5,000 Chinese characters, proactively ask AI to archive the memory: Write the key information from this conversation into MEMORY.md, then start a new session. 3\. Replace real-time processing with batch processing Do not process every message the moment it arrives Process messages in a batch once every 30 minutes Reduce the number of API calls V. Running the Numbers: Is Raising AI Worth It? 5.1 My Monthly Cost Breakdown Fixed costs: Cloud server (Alibaba Cloud ECS): RMB 200/month Domain + CDN: RMB 50/month Skill-market subscriptions: $15 (approximately RMB 105) Variable costs: Token consumption: $400–600 (approximately RMB 2,800–4,200) At a midrange estimate of $500: RMB 3,500/month Total Approximately RMB 3,850/month _Editorial arithmetic note: The displayed 200 + 50 + 105 + 3,500 adds to RMB 3,855. The source’s approximate RMB 3,850 total is retained. At the displayed alternatives, 3,850/600 is about 6.42, and the savings against RMB 8,000–10,500 are about 52%–63%; the following rounded “6 times” and “50–60%” remain the original wording, not exact calculations._ 5.2 Comparison with Alternatives Comparison fields: option; monthly cost; notes. Option A: hire a human assistant; monthly cost: RMB 8,000–10,500; includes salary, social insurance and a workstation. Option B: use SaaS services; monthly cost: approximately RMB 600; ChatGPT Plus + Claude Pro + Feishu tools. Option C: OpenClaw; monthly cost: RMB 3,850; autonomous and controllable, with a claimed 10-fold broader range of capabilities. Conclusion: 50–60% cheaper than a human assistant 6 times more expensive than SaaS services, but with a 10-fold broader capability range 5.3 Whether It Is Worthwhile Depends on How You Use It If your OpenClaw merely answers a few questions occasionally, it is not worthwhile. If your OpenClaw helps you every day to: Automatically generate morning and evening briefings Manage 3 Feishu groups Process 50+ messages Write 5 articles Monitor competitor developments Organize meeting minutes then it is more cost-effective than hiring 3 people. VI. Future Cost Trends 6.1 Will Token Prices Fall? Short term, within 1 year: The price war among Chinese models—Kimi, GLM, and MiniMax—will continue, with another estimated 30–50% of room for price reductions. Long term, within 3–5 years: As Token consumption becomes an infrastructure cost, prices will approach the cost of computing power and marginal benefits will diminish. 6.2 The “Metabolic Chain” vs. the “Intelligence Chain” 5Y Capital proposed a concept: the split into two chains. Comparison fields: dimension; intelligence chain; metabolic chain. Dimension: scenario; intelligence chain: a person sits at a computer collaborating with AI; metabolic chain: an agent operates autonomously in the background. Dimension: decision variable; intelligence chain: absolute intelligence; metabolic chain: absolute efficiency. Dimension: user loyalty; intelligence chain: extremely high and relatively insensitive to price; metabolic chain: low, but consumption is hundreds or thousands of times greater than in the intelligence chain. Dimension: logic; intelligence chain: luxury-goods logic—you would not buy a less capable copilot; metabolic chain: food logic—a living organism eats continuously, and economics determines survival. Conclusion: If you are an “intelligence chain” user, use the most powerful model and do not worry much about cost. If you are a “metabolic chain” user, Token costs are your lifeline. Most OpenClaw users belong to the metabolic chain. 6.3 Saving Money Is Not the Goal; Efficiency Is “ You must waste TOKENS. You must find every possible way to waste TOKENS. Only by wasting them can you explore the boundaries. — Manye (慢页), “Advanced Techniques for Asking OpenClaw Questions” This sentence does not mean that you should spend recklessly. It means: Do not limit AI’s capabilities merely to save money Treat Tokens as “energy,” not “cost” The ROI of exploration is far higher than the ROI of saving The ultimate purpose of saving money is to invest what you save in higher-value exploration. Final Thoughts After deploying OpenClaw for one month, I ran the numbers: Token consumption: $600 Cloud server: RMB 200 Total cost: approximately RMB 4,500 But I used it to complete: An automated morning-and-evening briefing system Intelligent management of 3 Feishu groups A competitor-monitoring system An internal knowledge-base question-answering bot An automatic code-generation tool Hiring people to complete this work would require at least 2 full-time employees and cost RMB 15,000+/month. Conclusion OpenClaw is not free, but it is the most cost-effective “employee” I have ever used. Of course, this assumes that: You know how to configure tiered models You know how to control Token consumption You raise it as a “living organism” rather than use it as a “tool” “ The greatest investment opportunity is not to build a better Agent, but to become the water and air that sustain the Agent world. — Investor at 5Y Capital In my view, for ordinary users: “ The greatest lesson is not how to save money, but how to find the balance between cost and capability. Free AI will not think for you, but AI you can afford can let you think less. Appendix: My Cost-Configuration Template { "budget": { "monthly\_limit": 500, "currency": "USD", "alert\_channels": \["feishu", "email"\], "alert\_thresholds": \[0.5, 0.8, 0.95\] }, "model\_tier": { "tier1": { "model": "claude-opus-4", "use\_for": \["complex\_reasoning", "architecture\_design"\], "max\_calls\_per\_day": 50 }, "tier2": { "model": "kimi-coding/k2p5", "use\_for": \["coding", "writing", "analysis"\], "default": true }, "tier3": { "model": "minimax/abab6.5", "use\_for": \["simple\_qa", "summarization"\] }, "tier4": { "model": "ollama/llama3.1", "use\_for": \["heartbeat", "monitoring"\], "local": true } }, "optimization": { "compact\_context": true, "archive\_threshold": 5000, "batch\_processing": true, "batch\_interval": "30m" } } Original source character-count note: approximately 4,500 Chinese characters Unique Research Data sources: real usage experience + 5Y Capital investment notes + a closed-door SiliconFlow presentation Disclaimer: Token prices fluctuate with the market; refer to actual bills --- Original publication: https://uniqueresearch.substack.com/p/src-20260214-01html On-site reading page: https://ffcap.cn/en/research/src-20260214-01html