The Repricing of Code: Zhipu GLM's Credit System and the Quiet Death of Subsidized Intelligence

The most consequential monetary reform of this quarter was not announced by the Federal Reserve, nor by the People's Bank of China, nor by any of the digital currency committees with which I have spent the past years consulting in Doha. It was quietly published on a pricing page for a Chinese AI coding assistant, a page most Western analysts will never scroll past, a page that nonetheless contains the same architecture of rationing, incentive, and scarcity that central banks spend entire careers constructing. Zhipu AI has taken its GLM Coding Plan, abandoned its old prompt-count limits, and replaced them with a credit-based accounting system that raises the monthly fee for new users by as much as 261 percent. On its surface, this is a commercial adjustment. Beneath it, tracing the liquidity ghost in the machine, it is a reallocation of scarce compute, a redefinition of what an AI interaction is worth, and the clearest signal yet that the era of subsidized intelligence is ending. The details, as confirmed by the emerging report, are stark. Under the old V2 scheme, Zhipu charged 49, 149, and 469 RMB per month for its Lite, Pro, and Max tiers, and supply was constrained by an unusual mechanical valve: every day at 10 a.m., a limited number of registration slots were released, and developers who arrived late found themselves queuing for a product they were willing to pay for but could not access. This is not pricing; this is rationing, and it should be understood in the same vocabulary we reserve for capital controls and foreign-exchange reserves. A queue is a ledger of unsatisfied demand. The new system replaces that queue with something more sophisticated: a unified credit ledger that meters four primary resource categories — input tokens, output tokens, cached context tokens, and MCP (Model Context Protocol) tool calls — and translates them all into a single synthetic unit of account. The headline numbers deserve emphasis. The new price points are 118, 538, and 1078 RMB for Lite, Pro, and Max, corresponding to increases of 141 percent, 261 percent, and 130 percent respectively. The Pro tier, the developer's default, absorbed the heaviest increase, which is the first sign that Zhipu's data science team knows precisely where its paying users cluster. Meanwhile, the grandfathering provisions are elaborate: existing V2 users may renew or upgrade at the old price, and V1 users have been granted a window before their plan expires to purchase at the old V2 pricing, with the purchase portal reportedly scheduled for mid-August. This combination — new customers priced high, old customers sealed in amber, a deadline that manufactures urgency — is not a random marketing mix. It is a structured migration campaign, and it tells us something important about where Zhipu believes the market is heading. The technical reading of this transition is where the story stops being a piece of Chinese AI news and becomes a case study in infrastructure economics. I have spent two decades auditing cryptographic systems and, more recently, modeling how central banks allocate settlement liquidity, so the accounting architecture here is the first thing I notice. The choice of metering units is revealing. The old model counted prompt numbers — requests per five hours, prompts per week. That is the monetary-policy equivalent of measuring money supply by counting banknotes, blind to credit expansion. It cannot distinguish between a single-token query and a multi-thousand-token codebase analysis. The new model can. By separating input from output, by isolating cached tokens as their own line item, and by folding MCP calls into the same ledger, Zhipu has created something that resembles a modern settlement system: a unified unit of account for heterogeneous compute resources, each with its own cost of production, each convertible into the same credit numeraire. This is the financialization of inference, and it represents the first time a major Chinese AI vendor has exposed its internal cost structure through the medium of a price list. The cache token line item deserves particular attention. From my audit experience, the explicit separation of cached context from fresh input is a subsidy in disguise. Cached tokens are dramatically cheaper to serve; the model does not need to re-embed and re-process the entire conversation history when the relevant context has already been computed. By itemizing cache tokens and, presumably, pricing them below fresh tokens, Zhipu is effectively paying developers to behave in ways that reduce its own inference burn. The pricing page is an incentive schedule disguised as a menu. This is the kind of fine-grained resource steering one normally encounters in the settlement layers of wholesale CBDC projects, not in a consumer coding tool, and its appearance here suggests that the engineering team behind GLM thinks like a monetary operator, not a feature vendor. Then there is the MCP metering, the detail most casual commentary will skip, and arguably the most significant single fact in the entire announcement. By billing external tool calls as a metered resource, Zhipu has functionally crossed the threshold from a conversational completion engine into an agentic execution platform. Every time the model reaches into a plugin, a repository, a cloud service, or an external data source, that interaction becomes a payable event. This is the same philosophical shift Ethereum underwent during the Merge. Proof-of-work charged validators in physical electricity; proof-of-stake charges them in financial commitment. In the GLM context, the old prompt-count limit charged for conversational surface area, while the new credit system charges for operational footprint. History rhymes in the ledger. The comparative arithmetic is worth laying out with care, because this is where a macro picture begins to emerge. GitHub Copilot sits near ten dollars per month, roughly seventy RMB. Cursor Pro sits near twenty dollars, roughly one hundred and forty RMB. Zhipu's Pro tier, at 538 RMB, is about 3.8 times the price of Cursor Pro and approaches eight times the price of GitHub Copilot. Even the Lite tier, at 118 RMB, costs more than Copilot's standard plan after conversion. This price point fractures the long-held assumption that Chinese AI products are natural substitutes competing on price. Zhipu is no longer selling a cheap alternative; it is selling a premium instrument, and it is doing so before publishing any benchmark comparison that would justify the premium. That ordering — price first, proof later — is a bold bet on brand inertia. The risk structure of this transition is equally instructive. The first exposure is churn: a 261 percent increase on the most popular tier invites a migration of price-sensitive developers toward Tongyi Lingma, CodeGeeX, or the international tools, and the switching cost for developers is lower than Zhipu might hope — context, habits, and plugin ecosystems can be rebuilt; model loyalty cannot be inherited. The second exposure is opacity: a credit system that consumes tokens invisibly breeds suspicion of what users will call "credit theft," and without a transparent consumption dashboard or published per-task conversion examples, the trust deficit will compound. The third exposure is the grandfathering cliff: when the protected V1 and V2 cohorts eventually migrate to the new pricing, the psychological jolt of paying multiples more for the same service can produce a renewal gap that the initial revenue bump will mask. The price list, in other words, is also a risk register. For those of us who watch liquidity rather than headlines, the deeper signal is what this reprice says about compute scarcity. A national champion of Chinese AI has effectively admitted, through its own pricing action, that its inference capacity is constrained. The daily 10 a.m. slot release was not a growth tactic; it was a throttle, and a throttle does not get removed by generosity. It gets replaced by a finer instrument, and price is the finest filter ever invented. Every price increase at the centralized layer is simultaneously an acknowledgment that marginal compute is expensive and a quiet confirmation that alternative architectures, where compute is tokenized and traded as a commodity on decentralized networks, gain credibility with each adjustment. The DePIN thesis that has circulated through crypto markets for two years is not validated by this single event, but it is fed by it. The dual-track pricing strategy deserves its own reading. Locking V2 users at old prices while opening a time-boxed window for V1 users to acquire legacy rates is functionally equivalent to an options structure. Legacy-priced users hold a call option on the old economic regime; the mid-August deadline is an expiry date; the controlled scarcity of the migration window is a synthetic volatility surface designed to trigger early exercise. I have seen this architecture before, in the coupon-stripping operations of sovereign debt markets and in the grandfathering clauses of CBDC pilot programs. It always serves the same purpose: to extract maximum settlement value from a transitioning user base while minimizing the optical damage of the transition itself. The standard reading of this reprice is cost-push inflation: inference is expensive, GPU supply is tight, and Zhipu must pass through costs to survive. Having built audit frameworks for exactly these transitions, I find that reading incomplete. The more likely objective is the construction of a revenue narrative for a forthcoming financing round or a strategic re-rating. Raising prices not because costs have risen but because monthly recurring revenue metrics need to look a certain way is a well-trodden playbook in private markets, and the evidence is in the details: the announcement discloses no capability improvement, no new model version, no benchmark table, no context-length expansion. When a price increase travels with zero disclosed value enhancement, it is an extraction event, not a value event. The blind spot runs deeper than the financial story. The credit system introduces an opacity that the prompt-count model never possessed. A developer can count prompts in five hours; a developer cannot easily audit how many cache tokens a conversation consumed, or whether an MCP call was billed at the expected rate, or whether the meter itself reports accurately. This asymmetry is where trust begins to erode — not through malicious design, but through the mundane friction of an unverifiable instrument. And in the privacy dimension, the persistence of cached context and the cross-boundary nature of MCP calls mean user code now rests on servers in configurations the old model never required. We sleepwalk into a digital panopticon one innocuous pricing page at a time; the developer who accepts cache metering has accepted, without a single checkbox, a new data-retention regime. Privacy eroded not by code, but by consensus — by the quiet agreement that convenience justifies a meter on everything. The mid-August window will close, and the queue may or may not reappear. Watch whether competitors undercut, whether Zhipu publishes a consumption dashboard, whether the SWE-bench rankings move. But the deeper signal is already legible: when a leading AI vendor rations access through price, intelligence has become a scarce asset, and in a bull market, scarcity flows into the ledger. History rhymes in the ledger. The question is no longer whether machine intelligence will be metered, denominated, and settled like every other liquidity. The question is who gets to hold the meter.