# A Modest Proposal for Anthropic > Source: > Published: 2026-08-31 05:54:22+00:00 # A Modest Proposal for Anthropic *The immediate occasion: @ClaudeDevs announced that today’s temporary 50% increase will become a permanent 25% increase on September 14—then acknowledged that, compared with what customers receive today, this is a 17% reduction. The account promised forthcoming changes intended to make customers “feel like” they are getting more. Original announcement; announcement and follow-ups.* ## For Relieving Anthropic of the Burden of Its Paying Customers, and for Making Their Unfinished Work Beneficial to the Company It is a melancholy object to behold the terminals, offices, bedrooms, and coffee-houses of this flourishing technological age crowded with developers who, having paid their monthly tribute to Anthropic, sit idle before a message informing them that they may resume their livelihoods on Thursday. These unfortunate persons may easily be distinguished. Their eyes are fixed upon the `/usage` command; their hands tremble over unfinished pull requests; and their mouths repeat those pitiable calculations by which they hope to discover whether a limit described as increased has, by some vulgar corruption of arithmetic, become smaller than the limit they possessed yesterday. Every person of sound commercial judgment must agree that this presents a grievous burden upon Anthropic. A subscriber pays a fixed monthly sum and then, with astonishing impudence, attempts to use the product. Some employ it daily. Others send long prompts. A few depraved persons even ask Claude to write code, though this is precisely the sort of behavior most likely to consume the compute which Anthropic had naturally hoped to preserve after accepting payment for it. The company has already pursued several ingenious remedies. Limits have been introduced, adjusted, divided into hourly and weekly varieties, temporarily enlarged, extended, partly withdrawn, reset, and renamed. Yet the disorder continues, because customers persist in comparing the service available today with the service available yesterday, rather than with a smaller historical baseline selected for the convenience of the announcement. I therefore offer the following proposal, which I hope will prove equally agreeable to Anthropic’s accountants, safety researchers, communications department, and suppliers of congratulatory launch tweets. Anthropic should abolish included usage altogether. Every paid subscriber should instead receive one ceremonial token per billing cycle. This token would not be sufficient to obtain an answer, but it might be used to ask Claude how much additional credit would be necessary to obtain one. The answer, naturally, would require the purchase of additional credit. This arrangement preserves every essential quality of a subscription. The customer continues to subscribe. The company continues to collect the subscription. Claude remains visible in the interface. Only the expensive and troublesome portion—the service—need be removed. The plan should be announced as the largest usage increase in company history. This may seem difficult to persons unfamiliar with modern product arithmetic, but the operation is simple. First establish a new standard allowance of zero. Then grant customers one token. This constitutes an infinite percentage increase over the standard allowance, a generosity so vast that no critic could express it numerically. Should any malicious person observe that the customer formerly received several million tokens, Anthropic may explain that those were promotional tokens, experimental tokens, unsustainable tokens, legacy tokens, or tokens whose existence was never intended to create an expectation that a paid product might remain usable. Some may object that customers will feel they are receiving less. This difficulty has happily been anticipated. Anthropic has promised improvements that will make users *feel* they are getting more, and feeling is incomparably cheaper to serve than inference. I propose, therefore, a new Usage Experience Dashboard. It shall contain a large green circle labelled **VALUE**, which remains permanently full. Beneath it, a smaller and less conspicuous meter shall show that no further messages may be sent until the heat death of the universe. Customers may be given complete control over the color of this meter, thereby fulfilling the promise of greater visibility and control. The five-hour limit should likewise be renamed the Five-Hour Opportunity Window. Upon exhausting it, the customer should receive the following encouragement: Congratulations! You have completed this period’s sustainable intelligence allocation. Your next opportunity begins at 3:00 a.m. The word *limit* possesses a disagreeable air of deprivation. *Opportunity*, by contrast, suggests that the customer has gained something—specifically, an opportunity to sleep, reconsider his profession, or open an account with a competitor. The economic benefits are plain. Suppose ten thousand Max subscribers each pay two hundred dollars per month. If they are permitted to use Claude, Anthropic must purchase electricity, operate data centers, maintain inference capacity, and answer support complaints about why Claude stopped in the middle of deleting a production database. But if none are permitted to use Claude, nearly the entire two million dollars may be preserved. Service reliability will approach one hundred percent, since a request that cannot be made cannot fail. Latency will fall to zero, measured from rejection to disappointment. Token efficiency will become perfect. No model in history has produced fewer unnecessary words than one which produces none. There will also be considerable savings in safety. Anthropic has long labored to build an artificial intelligence incapable of assisting with dangerous acts. Yet even the most carefully aligned model may occasionally answer a question. Under my proposal, this residual risk is abolished. A model without available quota cannot give instructions for wrongdoing, introduce a security vulnerability, hallucinate a citation, offend a user, or embarrass the company in a benchmark comparison. It would be the first perfectly safe general intelligence, because nobody could consult it. The proposal also offers a humane solution for developers whose work is interrupted by the limit. At present, their unfinished code lies uselessly upon their machines. Anthropic should claim these fragments as abandoned computational property. They may be used for training, sold as synthetic data, or bundled into a benchmark entitled *Real-World Agentic Tasks Which Claude Nearly Completed Before Thursday*. The developer ought not to complain. He has received the educational benefit of beginning the work, while Anthropic receives the commercial benefit of owning its remains. Thus the same project feeds both personal growth and model improvement, although admittedly in different proportions. Further revenue may be obtained from the reset itself. At present, limits reset automatically, a practice that leaves money scandalously uncollected. I recommend the following products: - Reset Preview, permitting the customer to see when usage might return. - Reset Plus, permitting the clock actually to reach that time. - Reset Max, ensuring that the new allowance exceeds the tokens consumed while asking why it did not reset. - Reset Enterprise, in which a sales representative explains that no one is authorized to disclose the limit. The `/usage` command should consume usage. This will encourage customers not to inspect their allowances too often, thereby increasing the probability that they mistake scarcity for abundance. Likewise, every complaint posted on social media should automatically spend five percent of the complainant’s weekly quota. This is equitable, since public discussion of Claude constitutes Claude-related activity. Praise, being less computationally expensive to moderate, may receive a two-percent bonus. I can think of but one objection possessing the slightest appearance of weight: that some customers may cancel. This, however, is not a defect but an elegant mechanism of capacity management. Every departing subscriber reduces demand. Anthropic may then announce that availability has improved for remaining users, whose limits can consequently be reduced to protect the newly improved availability. If a sufficient number depart, the company will at last possess ample compute for every customer it no longer has. Others may contend that a fixed-price subscription ought to disclose a reasonably stable quantity of service. But such persons mistake the purpose of a subscription. A subscription does not purchase usage; it purchases a recurring relationship with a billing system. The amount charged is fixed because customers value predictability. The amount delivered must remain variable because Anthropic values flexibility. A few sentimentalists may recommend the opposite course: publish exact token-equivalent allowances, give adequate notice of reductions, maintain the purchased terms until renewal, or charge transparently by consumption. These ideas are manifestly ruinous. Exact numbers might be compared. Adequate notice might permit planning. Stable terms might create trust, which is expensive to acquire and impossible to list as infrastructure capacity. Above all, plain language would deprive the communications department of its most important innovation: the conversion of subtraction into growth. Consider the beauty of the present formula. An allowance rises from an ancient baseline of one hundred to a temporary one hundred and fifty. It then descends to one hundred and twenty-five. The customer, who foolishly measures from the service in his hands, calls this a reduction. The company, measuring from a historical quantity selected for the purpose, calls it an increase. Both statements are mathematically defensible, but only one is useful for marketing. My proposal merely completes this science. Every reduction should be announced relative to a still lower imaginary number. If weekly usage falls from one hundred tokens to ten, let the standard baseline be declared one token, and announce a permanent nine-hundred-percent increase. If it falls from ten to none, compare it with negative ten and celebrate a hundred-percent improvement. In time, Anthropic may furnish no inference whatever while announcing increases of unprecedented magnitude. The social advantages are no less considerable. Programmers deprived of Claude will rediscover documentation. Managers will rediscover schedules. Venture capitalists will discover that a demonstration is not a business. Humanity, being unable to ask an artificial intelligence to perform every task, may briefly remember how to perform one itself. Meanwhile, Anthropic will retain the subscriptions, the compute, the praise for increasing limits, and the moral distinction of having prevented its technology from being overused by precisely those persons who paid to use it. I profess, with complete sincerity, that I have no private interest in advancing this necessary work. I possess neither data center nor GPU; I hold no shares from which I might profit; and my own weekly allowance is already exhausted. It resets on Thursday. *Written by Tornike Onoprishvili, partially with GPT-5.6-Sol.*