Editor's note: This is a complete translation of the original February 21, 2026 commentary. Prices, valuations, usage reports and investment opinions are retained as historical statements by the source and its quoted developers, not independently verified current figures. The original calls February 20 “two days later” after mentioning February 12; those dates are eight days apart. Its token-to-character conversion and illustrative book comparisons are not internally consistent. The quoted startup cost comparison of RMB 5,000 versus RMB 120,000 implies about 95.8% savings, rather than the stated 80%. These source statements are preserved below so the translation does not silently rewrite the record. The 700-fold valuation comparison uses the source’s nine-month revenue and reported market capitalization; its currency conversion, annualization basis and comparability with OpenAI are not independently established. Personal usage reports, trading-profit claims, model/plan labels and statements about pricing power remain attributed anecdotes and opinions, not forecasts validated by this edition.
Original · Unique Research · 2026-02-21
COVER STORY
I Ran the Numbers: In the Agent Era, My Token Bill Will Surge 100-Fold—What Justifies Zhipu AI and MiniMax Trading at a Price-to-Sales Ratio of 700?
Last night, I opened Zhipu AI’s dashboard and saw that this month’s Coding Plan had already consumed nearly 190 million Tokens. When the bill appeared, I stared in shock for three seconds.
Before the 30% price increase, I used at most 80 million a month. Now, with GLM-5 released, a single complex Agent task consumes tens of millions, and I still somehow think it is “worth it.”
I am not the only one going mad. On February 12, the same day Zhipu AI issued its price-increase notice, the new Coding Plan packages sold out immediately upon launch. MiniMax released its M2.5 Agent-native model that same day, and developers instantly bought out the Highspeed version as well.
Two days later, on the first Hong Kong trading day of the Year of the Horse—February 20—the Hang Seng Tech Index fell nearly 3%, yet Zhipu AI surged 42.72% to close at HKD 725, sending its market capitalization above HKD 323.2 billion. MiniMax rose 14.52%, and its market capitalization also surpassed HKD 300 billion.
The valuation stands at a price-to-sales ratio of 700. MiniMax generated only $53.44 million, approximately RMB 376 million, in revenue during the first 9 months of 2025. At its current market capitalization, its price-to-sales ratio exceeds 700—an order of magnitude higher than OpenAI’s 65 times.
What exactly is the market betting on? Once I calculated my Token bill, I understood.
My Billing Story: From RMB 200 to RMB 1,000, and I Think It Is Worth It
First, let me explain how I use Zhipu AI.
I started in October 2024. GLM-4 had just been released, and my main goal was to replace GitHub Copilot. Copilot cost $10 a month, while Zhipu AI offered a first-purchase discount at the time and cost only a few dozen yuan per month. It was an excellent deal.
The first few months were stable. I wrote code, debugged software, and wrote documents every day, using around 80 million Tokens per month at a cost of RMB 200–300. I thought it was economical.
The turning point came in January this year. I began combining OpenClaw with Agents so AI could automatically handle repetitive work—collecting competitor information, generating weekly reports, and answering standardized questions. With 1 Agent running, Token consumption rose to 150 million per month. I still found that acceptable because it saved me considerable time.
But once GLM-5 launched, the situation changed completely.
I began trying more complex Agent tasks: multiple Agents developing projects in parallel, Agents automatically analyzing codebases and refactoring code, and Agents automatically writing technical documentation and generating test cases. The bill exploded. In the first 20 days of February, consumption reached 190 million Tokens; projected full-month consumption was 250–300 million, costing RMB 800–1,000 after the price increase.
The problem was that I still somehow thought it was “worth it.”
That was because this month I used AI to complete work that previously required 3 people: refactoring code for 2 complete projects, writing 15 technical documents, producing 20 competitor-monitoring reports, and automatically managing 3 groups with an average of more than 200 daily messages. Hiring people to do this would cost at least RMB 15,000 per month.
100-Fold Consumption Is Not an Exaggeration; It Is a Conservative Estimate
The “100-fold” increase I describe is not my story alone.
Consider what is actually happening to other developers.
One backend engineer previously wrote code and searched documentation manually, consuming 500,000 Tokens a day. Now he has configured 5 Agents for architecture design, code implementation, testing, documentation, and operations. He averages 6.3 million Tokens per day and peaked at 97 million Tokens. In his words: “I used to write code. Now I watch an Agent write code and then review it. Efficiency has improved, but the Token consumption really is frightening.”
At another AI startup, a 10-person company configured 30 Agents: some answer customer-service questions automatically, some analyze data, and some generate content. The company consumes 1.5 billion Tokens per month at a cost of RMB 5,000. Its CTO calculated: “Hiring people to complete this work would require at least 15 people and cost RMB 120,000 per month. The Agent approach saves 80%.”
The most extreme case was an independent developer running an automated-trading Agent that monitors markets, analyzes data, and executes trades 24/7. Daily Token consumption peaked at 120 million, and monthly costs exceeded RMB 4,000. His feedback was: “At first, spending Tokens hurt. Then I ran the numbers—the money my Agent earns is enough to pay 100 years of Token bills.”
These are not isolated cases. An ordinary programmer can write only several thousand lines of code and read tens of thousands of Chinese characters of documentation per day. An Agent can write code, invoke tools, reflect, iterate, and perform multiple rounds 24/7, easily consuming 10 million to tens of millions of Tokens a day.
What does 10 million Tokens mean? In Chinese, 1 Token is approximately 1–1.5 Chinese characters, so 10 million Tokens equal 7–10 million characters. That is equivalent to reading 12–18 copies of “War and Peace,” reading the entire “Harry Potter” series 7–9 times, writing a 300,000–500,000-character novel plus all its annotations and iterative versions, or producing 3–5 years of a programmer’s code.
A person can read at most 100,000 Chinese characters and write 10,000 in a day. An Agent can work for 24 hours without sleep and multiply your daily workload by 100. What happens when 10 or 100 Agents run behind one person in the future? Token consumption explodes exponentially.
That is why I say “100-fold”—not as an exaggeration, but as a conservative estimate.
How Can Zhipu AI Raise Prices by 30% and Still Sell Out?
A CICC research report put it clearly: the industry’s pricing logic is shifting from “traffic consumption” to “monetizing the value of computing power.”
Recall the first half of 2024, when the large-model price war was brutal: ByteDance’s Doubao cost RMB 0.0008 per thousand Tokens, Alibaba’s Qwen model at the GPT-4 level cut prices by 97%, and Zhipu AI itself had cut prices by 90%. The logic then was to seize market share and grow volume even at a loss.
What about now? Zhipu AI raised Coding Plan prices by 30%, removed the first-purchase discount, and doubled overseas API prices. The new packages nevertheless sold out immediately and added weekly limits. MiniMax M2.5 Highspeed runs at 100 tokens/s and costs only $1 for 1 continuous hour, yet developers bought out the high-speed version.
This is not exploiting customers; it is a signal that pricing power has genuinely become established.
Why? Because Tokens are becoming the “new oil.”
Oil’s value lies not in oil itself, but in its ability to power cars, aircraft, and factories. A Token’s value lies not in the Token itself, but in its ability to drive Agents to complete work. When models become capable enough and Agents begin deploying at scale, the supply-demand relationship for Tokens will fundamentally reverse.
Supply is linear—expanding GPU capacity, power, and chips takes 1–2 years. Demand is exponential—every improvement in a model unlocks 10 times as many new scenarios. This is why Zhipu AI can raise prices by 30% and still sell out. Users are not buying “software”; they are buying “energy that burns continuously.”
The Developers’ “Worth Every Penny” Moment
I collected authentic feedback from developers.
One programmer said: “The Pro package is enough for me, but I renewed Max anyway because only GLM-5 can complete complex distributed-systems engineering in one attempt. I used to spend more than $200 a month on GPT; now Zhipu AI Max costs RMB 800 a month, which is still acceptable.”
Another AI entrepreneur said: “I used to worry about wasting Tokens on throwaway code. Now I can submit anything to GLM Coding Plan, and efficiency takes off. The time from conceiving a feature to launching it fell from 3 days to 3 hours.”
Ajie, an independent developer, was more direct: “GLM-5 matches Claude Opus 4.5 and is even faster. After the price increase, I did not consider switching to a competitor. Honestly, as long as quality remains stable, I would accept another 30% increase.”
An algorithm engineer at a major technology company told me: “Our team migrated from OpenAI to Zhipu AI, cutting costs by 40% with roughly comparable results. New internal projects now use Zhipu AI by default, and legacy projects still using OpenAI are gradually switching.”
These voices are not marketing; they reflect genuine user stickiness. The market is voting through a price-to-sales ratio of 700: Tokens will become an infrastructure-level consumable for every person and every enterprise, like electricity. You do not value a power company on PS; you look at utilization, pricing power, and the demand curve.
But I Must State the Risks Sharply
The logic is sound, but the “window of time” is the real bet.
MiniMax’s prospectus shows that revenue grew 170% in the first 9 months, yet its net loss reached $512 million—losing 10 units of money for every 1 unit of sales. Although losses are narrowing, breakeven remains distant.
More than 70% of MiniMax’s revenue comes from overseas, while the long-term user stickiness of its core product, Talkie, has not yet been proven. Zhipu AI’s Coding Plan is popular, but how high is the ceiling for coding assistants? When Claude, Gemini, and GPT all compete in the same market, how long can pricing power last?
Another easily overlooked factor is that this surge occurred in Hong Kong, where pure-play AI stocks are extremely scarce: only Zhipu AI, MiniMax, and Haizhi Technology. When large amounts of capital flow into an extremely narrow sector, valuation premiums are amplified sharply. This is not driven entirely by fundamentals; the structure of capital also matters.
The valuation at a price-to-sales ratio of 700 is betting that “the Agent era will erupt fully next year.” If it arrives 3–5 years later, today’s pricing may not hold.
But one point is already irreversible. When a 30% price increase still sells out and developers prefer paying more to losing access, supply and demand have completely reversed. The dawn of Token economics has arrived; the only questions are how long the dawn will last and when the sun will truly rise.
The Three Actions I Took After Running the Numbers
First, I renewed the annual Coding Plan to lock in a relatively low price.
Second, I began researching how to move simple tasks to cheaper open-source or tiered models, reserving top-tier Agents for Zhipu AI and MiniMax. The specific strategy is to use MiniMax M2.5 Highspeed or a local model for simple tasks such as information collection and format conversion; GLM-5 Pro for medium tasks such as code generation and document writing; and GLM-5 Max or Claude Opus for complex tasks such as architecture design and complex reasoning.
Third, I reassessed my Agent configuration. Not every task needs a top-tier model: monitoring Agents can run 24/7 on inexpensive models; creative Agents can use top-tier models when needed; and interactive Agents can use mid-tier models to balance cost and quality.
We Have Only Just Begun Learning to “Use Electricity”
In 2000, when China had 80 million mobile users, who could have imagined that the mobile internet would rewrite everything? Today, Tokens stand at a similar inflection point.
How much did your Token bill rise this month? How do you plan to handle a 100-fold “electricity bill” in the future?
Share your real usage and strategy in the comments. Together, let us witness—and prepare for—the historic moment when Tokens become the “electricity” of a new era.
Data sources: real usage experience + developer-group research + public financial reports
Disclaimer: investing involves risk; this article does not constitute investment advice