GLM Coding Plan’s Credit Flip Is a Hidden Ledger — and a Compute Rationing Signal

Ten a.m. was the signal. That was the moment the old GLM Coding Plan handed out its daily quota. Developers treated it like a treasure drop: refresh, try to grab a slot, fail, repeat. Then Zhipu pulled the slot machine and put in a meter. New users now pay 118, 538, or 1,078 RMB per month. Old V2 users keep their old rates. Old V1 users get one final discount window before mid-August. This is not an AI subscription story. It is a fee-market event with a hidden order book. I have sat on both sides of this table. In 2017, I raised $4.2 million in 48 hours for a white-label ICO by selling a story about decentralized sovereignty. I did not audit the token mechanics, and I paid for it. By 2020, I was running a security review for AeroSwap, and I found a reentrancy vulnerability in the withdrawal function that would have burned $15 million in TVL. The lesson from both experiences is the same: when a system changes its unit of account, the real information is hidden in the pricing, not in the announcement. Zhipu just changed its unit of account. Let me show you what it is actually saying. GLM Coding Plan is Zhipu AI’s code assistant, built around the GLM series of large models. Under the old model, Zhipu throttled usage by prompt counts: a fixed number of prompts every five hours, and another ceiling every week. That is a coarse throttle. It counts conversations, not cost. It treats a one-line completion the same as a full repository rewrite. It makes every request look identical to the server, which is technically absurd. The new credit system fixes that absurdity. Input tokens, output tokens, cache tokens, and MCP calls all burn credits. Four resources, one ledger. Start with the numbers. The old V2 tiers were 49, 149, and 469 RMB per month. The new tiers are 118, 538, and 1,078 RMB. Lite is up 141%. Pro is up 261%. Max is up 130%. If you only read the top line, this is a price hike. If you read the gradient, it is a demand map. The first thing I noticed is the Pro tier. A 261% increase is not a rounding error. It is a statement about where the heavy users sit. The old pricing had Pro too close to Lite. The team knows that programming users cluster in the middle of the price stack. They are the ones consuming long context, calling external tools, and burning output tokens. Raising Pro by more than 2.6x while only raising Lite by 1.4x tells you where the real cost lies. Those users are also the most likely to churn if the credit burn rate feels dishonest. Now look at the resources. Input tokens are cheap. Output tokens are expensive. That is the basic economics of autoregressive language models: every generated token is a sequential computation, while input processing can be done in a broad parallel pass. The old prompt-count model could not see this. A user could ask one giant question with a 50,000-token context, get a short answer, and consume far less than a user who asked ten short questions and demanded ten generated code blocks. The credit system fixes the mismatch. This is the same granularity upgrade that decentralized protocols went through when they moved from simple transaction fees to calldata pricing and state access costs. The quietest detail in the announcement is that cache tokens have their own line item. In a ordinary subscription, caching is an implementation detail. In a credit-based system, it is a discount program. Zhipu is telling users: reuse long context and pay less. That is a direct admission that context retrieval is expensive. It is also a subtle shove toward what I call context engineering. We saw this in crypto with gas optimization: the protocol becomes cheaper when users compress their own transactions. Zhipu is now rewarding developers who design their prompts around cached prefixes. That is a behavioral shift, not just a billing tweak. MCP calls on the meter is the biggest tell. MCP, or Model Context Protocol, is how an AI model reaches outside its own weights. It reads files, queries databases, triggers tools, and calls other services. Billing MCP calls means Zhipu is no longer selling a chat box. It is selling an agent runtime. The credit system is the pricing layer for that runtime. This is what a platform transition looks like before it gets a slick marketing name. It also raises a security question I have seen before in DeFi: if a third-party tool can be invoked from inside the model, who audits the permission boundary? A credit meter does not solve that. It just makes the tool call visible on a bill. Then there is the two-tier migration strategy. The fact that V2 personal and team users can keep paying the old price is a retention play. The fact that V1 users get a chance to buy in at the old V2 price before mid-August is a conversion window. This is not a random schedule. It is a lockup schedule. Zhipu is deliberately creating a last-chance buying signal to compress old users into the new system before the Q3 numbers get refreshed. I saw this in 2017 ICOs: a sunset round that feels like a gift but is actually a demand extraction engine. Nobody says it directly, but the message is clear: get in at the old price or get left behind. The old version also had a daily quota that opened at 10 a.m. and sold out to “excessive demand.” Some journalists will call that a growth story. I call it a GPU P&L statement. Rationing only exists when marginal supply is expensive or fixed. If Zhipu had spare inference capacity, it would sell more low-price subscriptions and report a bigger user number. Instead, it chose to raise prices. That is exactly what a capacity-constrained supplier does. It is the same playbook as Ethereum’s EIP-1559: when demand exceeds base capacity, let the price float up until the queue disappears. Here is where the blockchain angle becomes impossible to ignore. Ethereum’s gas market is transparent. Every base fee is visible on-chain. Every user can see the price of block space before they submit a transaction. Zhipu’s credit system is not transparent. The conversion table between credits, tokens, and MCP calls is not public. The burn rate per typical programming task is not public. There is no dashboard showing why a code review consumes 45 credits and a chat session consumes 12. That asymmetry is the real risk. Let me put it in language I used during my LayerZero hackathon days. Zhipu is issuing points. Points are an unlisted token with a publisher-controlled exchange rate. The publisher can change the rate at any moment by altering the credit cost of a token or a tool call. If you cannot audit the burn rate, you cannot verify whether the 261% price increase comes from better model research or from a revenue target. The economic term for that is a black-box fee schedule. That is fine for a startup with pricing power. It is dangerous for a platform that asks developers to build mission-critical workflows on it. Now let me answer the obvious comparison. GitHub Copilot is about $10 per month. Cursor Pro is about $20 per month. GLM Coding Plan’s Pro tier is 538 RMB, roughly 3.8 times Cursor Pro. On a pure price-per-seat basis, Zhipu is now a premium product. But that comparison is too simple. Zhipu is not selling a generic code completion plugin. It is selling a vertically integrated runtime: Chinese-language code comprehension, local deployment options, enterprise data compliance, and a deeper connection to the GLM model family. For a developer in Shanghai who needs an AI assistant that understands Chinese comments, Chinese business logic, and Chinese regulatory patterns, Copilot at $10 a month is not competing on the same axis. The contrarian take is that this pricing move is actually bullish for decentralized AI infrastructure. Here is why. Centralized labs are discovering what miners learned in 2017: when demand outstrips supply, you do not add capacity immediately. You raise the base fee. That is a rational short-term move. But it also creates an opening for networks that provide verifiable compute pricing. If a decentralized inference market can show a public ledger of model execution costs, it does not need to ask developers to trust a point system. It can show the cost of every inference is a consequence of the market. Zhipu’s move is a signal that AI compute is becoming a priced, rationed, metered resource. That is the core thesis of many crypto projects. The centralized incumbents are proving the thesis while keeping the meter opaque. There is another hidden message in the price gradient. The Max tier at 1,078 RMB is a pricing anchor. It is not meant to have many buyers. It is meant to make the Pro tier at 538 RMB look reasonable. That is a classic behavioral pricing trick, but it matters for a different reason: it signals that Zhipu believes there is a segment of users for whom 1,078 RMB is acceptable. Those users are not freelance developers. They are engineering leads, team owners, or enterprise buyers who want a premium tier with priority GPU access. Zhipu is not trying to win the mass market. It is trying to win the high-value segment before competitors can establish trust. What should a developer do with this information? The answer depends on whether you can measure the credit burn. If you are an existing user, open the product and record your usage. Run one code review with a long context and note the credit balance. Run one agentic session with three MCP calls and note the balance. Build your own conversion table. Do not wait for Zhipu to publish one. If the burn rate is too high, the price increase is not a signal of model quality; it is a signal of cost recovery from the wrong customers. If the burn rate is low enough, the new price might still be rational because the credits are acting as a spending control mechanism for heavier workflows. For investors, this is a unit-economics upgrade. It lifts monthly recurring revenue without requiring new user acquisition. It demonstrates pricing power at a moment when AI valuation narratives need revenue quality. But it also creates a disclosure burden. If Zhipu does not publish a credit conversion table, the community will eventually suspect that the credits are a prepaid vault with hidden withdrawal fees. I audited a vault like that once. It ended in a run. The pricing move is logical. The opacity is the flaw. So what is the actual new information here? It is not the price. It is the meter. Zhipu has told the market that AI coding assistance is now a metered resource, with at least four distinct cost drivers: input, output, cache, and tool orchestration. That is a more precise technical admission than any benchmark. The company has also told us that its inference stack is capacity-constrained enough to warrant daily rationing and a 261% mid-tier price increase. And it has told us that it believes its model, data-compliance, and ecosystem position can survive a price point far above international rivals. Here is my 90-day checklist. First, does Zhipu publish a typical-task credit table? Show me that a 10,000-token code review costs X credits and a full agent session costs Y. If that table appears, the premium price has a rational floor. Second, does the V1 purchase window actually open in mid-August? That timing will confirm whether this is a planned migration or a reaction to utilization. Third, do independent reviewers run side-by-side cost comparisons against Copilot and Cursor? The first credible comparison chart will define the public narrative. If those tables do not appear, the credits remain a closed book. And a closed book is the one thing a technical community will not forgive. We did not have a name for this in 2017 because we were too busy raising money in 48 hours. Now we do. It is a hidden ledger. Do not buy the narrative. Read the meter.