One thousand seventy-eight yuan per month. That is the new ceiling on Zhipu AI's GLM Coding Plan, the product that once gave away its seats through a daily 10 a.m. lottery. The Max tier jumped from 469 yuan to 1078 yuan. The Pro tier went from 149 to 538. The Lite tier went from 49 to 118. In a single pricing update, one of China's most visible coding assistants mutated from a subsidized growth experiment into the most expensive developer subscription in the market โ roughly 3.8 times a Cursor Pro seat, 7.7 times GitHub Copilot, and more than many Chinese developers pay for their entire workstation cloud stack.
The market is not rational; it is resistant. This price tag is a wall.
But the headline percentage increases are the least interesting part of the change. What matters is the metering architecture. The old plan rationed by prompt count โ a five-hour window and a weekly cap that mapped neatly onto chat-style workflows while telling Zhipu almost nothing about compute consumed. The new plan is a credit economy. Input tokens, output tokens, cached tokens, and MCP calls each burn a specific category of credits. The message hidden under the price ladder is that Zhipu has re-plumbed its entire cost accounting.
I spent 2017 auditing more than fifty ICO whitepapers for a Stockholm venture fund, hunting for supply-chain vulnerabilities before token sales launched. The red flag I trusted above all others was the utility shift โ the moment a project changes its accounting unit from "requests" to "credits." It always means the team is rewriting the economics of scarcity. The public reads a pricing decision. Engineering teams read a production constraint.
The GLM credit ledger is no exception.
Context: From Prompt-Count to Credit
Let me lay out the migration sequence, because it reveals strategic architecture rather than a random adjustment. V2 personal and team users can continue on their original price. V1 users, who bought the earliest vintage of the plan, get a preserved purchase window at the old V2 rate. The new V1 purchase entry is expected to go live in mid-August. New users, meanwhile, face an entirely different rate card: Lite at 118 yuan, Pro at 538 yuan, Max at 1078 yuan.
This dual-track approach is textbook retention engineering. Grandfathering protects the installed base. The V1 window creates a scarcity event โ a final chance to buy at the old rate โ which historically spikes conversion. And the large spread between old and new prices makes on-time renewal feel like a reward rather than a decision. Zhipu is buying time against churn while simultaneously testing whether a new, more expensive customer segment will materialize.
The shift from prompt-count limits to a four-axis credit model reinforces the point. A prompt limit is a coarse throttle sized for chat sessions. A credit system with separate meters for input tokens, output tokens, cached tokens, and MCP calls is the accounting instrument of a platform that intends to monetize machine workloads, not just human conversations. That distinction matters because it defines the true competitive set: Zhipu is no longer selling autocomplete. It is selling compute capacity.
The Mechanical Truth of the Credit Architecture
The cached-token line item is the quiet tell. The old prompt-count regime charged one unit whether the model re-read an 80,000-token context or answered a trivial query. That is commercial indifference to the physics of inference. Autoregressive generation consumes energy for every input token and more for every output token. A repeated context read is a small furnace burning the same fuel again. By separating cached tokens, Zhipu has accomplished two goals at once. It has acknowledged that cache reuse is meaningful, separately measurable cost. And it is teaching developers, through price, to engineer longer-lived contexts and reuse them aggressively. In a world where a cached token is billed at a fraction of a fresh input token, the rational developer builds a persistent context layer. That is not billing. That is behavioral load-shedding.
Fractures in the ledger reveal the truth of value. The fracture between cached and uncached tokens contains the real cost model.
The MCP line item is an architectural confession. You do not build a four-way billing split โ input, output, cache, tool calls โ for a chatbot. You build it for an agentic orchestrator. The GLM Coding Plan has moved beyond completion. It is a system that reads repositories, invokes external tools, negotiates with APIs, and executes multi-step tasks. That is a different operational surface, and the credit system is the accounting skeleton for it. The plan is not competing with Copilot's autocomplete; it is competing with the entire agentic coding stack.
This is also where my crypto-native lens, working daily on tokenized compute markets like Render, sees an old pattern. Rating scarce resources through a credit system is a fee-market design. Blockchain networks do it with gas. Zhipu is doing it with points. Both systems triage demand by price, optimize for high-value requests, and reserve supply for the users who can pay. The difference is that Ethereum publishes its gas schedule. Zhipu has not published its credit equivalents.
It gets worse when you look at the cache-token incentive from the security side. A cache is a persistence surface. Every cached token is a snapshot of developer code, and that snapshot must live somewhere. The credit table quietly tells enterprises that cached memories are now a genuinely valuable โ and expected โ part of the production workflow. That is a wonderful optimization story for the vendor. It is a governance nightmare for a compliance officer who now has to ask where exactly the cache lives, how long it survives, what happens when a user leaves the plan, and whether the deletion is cryptographic or merely logical. Without those answers, the caching discount is priced in credits but paid for in risk.
The Back-of-the-Envelope Problem
Here is a simple arithmetic framework. Suppose the old Lite plan, at 49 yuan, gave a professional user roughly forty to sixty interactive coding queries per day inside the per-window and weekly caps. At 118 yuan, the new Lite plan is 2.4 times more expensive. The question is whether the initial credit grant is also 2.4 times the old effective allowance. If Lite ships with a generous grant and typical tasks cost modest amounts, the price increase is partially a transfer from conversation-count to token-volume pricing. If the grant is tight, the Lite tier becomes a demotion dressed as a price hike. The same logic applies to Pro and Max. Without published consumption examples, the pricing is undecidable. Fans will assume goodwill. Skeptics will assume extraction. That ambiguity is a risk, not a benefit.
I watched the same dynamic in the 2020 DeFi summer when liquidity farms promised infinite depth. The promise was always finite under stress. Zhipu's old plan was a finite promise; the new plan is a finite promise with a transparent pricing signal. The problem is that clarity in unit price without clarity in unit quantity is a half-open book. The vendor knows the credit burn rate per task. The buyer does not. That asymmetry is the source of almost all the coming controversy.
The cached-token design, to be fair, offers one bright counter-narrative. If a cached token costs one-tenth of a fresh input token, a developer who invests in context reuse can reduce effective task costs by forty to sixty percent. The credit architecture rewards software craftsmanship. Light users who treat the plan like a simple chat window may find the new economics brutal, while power users who adopt the cache, MCP, and batch patterns can make the numbers sing. The plan has quietly separated the market into skill classes. That is sophisticated incentive engineering โ but casual developers will experience it as a tax.
Security Is the Denominator Nobody Is Pricing
I cannot evaluate this change without a cybersecurity threat model because every component of the new ledger touches the three most sensitive layers of a developer's environment.
First, code. Any caching mechanism stores proprietary source code on vendor infrastructure. Cache persistence duration, access control, and deletion guarantees determine whether that data is safe. The Chinese legal environment has made it clear that code containing personal information, credentials, or algorithmic secrets falls under the Personal Information Protection Law and the Data Security Law audits. A cash-flush AI vendor can absorb a regulatory fine; a developer's leaked private key cannot be absorbed at all.
Second, credentials. MCP calls routinely involve authentication tokens. The credit ledger bills those calls, meaning the system records metadata about which third-party services a developer touches. This is not inherently evil, but it is a supply-chain observation surface that most consumer coding tools do not expose. The metering architecture gives Zhipu far more visibility into enterprise developer behavior than any previous generation of tools I evaluated during my audit years.
Third, supply chain. A compromised MCP adapter in an agentic coding plan is a remote code-execution vector. The protocol grants the model access to tools; the tools grant access to secrets; the secrets grant access to production systems. A malicious or poorly audited MCP service in the dependency graph is exactly the kind of vulnerability I would flag in a whitepaper review. My single piece of operational advice: do not connect production infrastructure to the Pro or Max tier until Zhipu issues a data-handling whitepaper that covers cache deletion, MCP permission boundaries, and invocation audit logs. The math may be sound; the risk is not yet measured.
The security question also doubles as a market-segmentation filter. Individual developers can tolerate opacity because their blast radius is small. Enterprises cannot. If Zhipu wants to convert the high ARPU of 538 yuan and 1078 yuan tiers into durable institutional revenue, it must produce compliance documentation that satisfies corporate security teams. The credit system creates the financial logic for enterprise adoption; the missing security whitepaper is the bottleneck on actual deployment.
The Retreat Beneath the Confidence
Every mainstream take reads this announcement as a confidence move. Pricing power. Premium positioning. A Chinese OpenAI moment. I read the exact opposite: a strategic retreat dressed as a premium excursion. A vendor with boundless capacity and unshakeable share does not need a 261% overnight increase on its most popular tier. It executes that move when the cost side is roaring and the demand side is already proven. The details prove the choreography. V2 users keep old prices. V1 users get a preserved purchase window at the old V2 rate. Mid-August entry is deliberately scheduled. This is a company preparing for a wave of cancellations and trying to bend the curve of attrition. It has chosen lost users over lost margins.
The deeper decoupling is between price and demonstrated capability. Previous AI pricing waves in both the United States and China were anchored in a benchmark โ a new model, a new reasoning score, a new context length. Zhipu's announcement carries no SWE-bench ranking, no HumanEval number, no third-party comparison. The credit ledger is presented as its own justification. Without an external quality anchor, the only contract between Zhipu and the developer is the conversion rate between credits and useful work. And that contract is incomplete.
Scale is a subsidy that expires at the worst moment. The old GLM pricing was exactly that โ a subsidy built on low costs and low expectations. Now the subsidy is gone, and the question is whether the capability stack can survive contact with a fully priced market.
That silence is a gift to every competitor in the sector. Alibaba's Tongyi Lingma, CodeGeeX, Baidu's Comate, and the international tools already priced below 1078 yuan have all inherited several hundred thousand suddenly-price-sensitive developers. The short-term carve-out belongs to the rival that publishes the cleanest comparison sheet and the most transparent credit calculation. Zhipu has not just raised its own price; it has lowered the cost of switching away from it.
The Door for Enterprises and the Wall for Casual Users
Enterprise customers complicate the doomsday story. A company purchasing ten thousand Pro seats at 538 yuan per seat is making a procurement decision with a different formula. Compared to the annual cost of a human pair programmer or a full-time machine-learning engineer, 538 yuan is negligible. For that buyer, the credit architecture is not a price increase; it is the skeleton of governance, cost allocation, and department-level chargeback. The consumer segment sees a wall. The enterprise segment sees a door.
And the macro ecosystem is listening. I track global liquidity as a daily discipline, and the current AI capex cycle is the largest synchronized reallocation of capital toward compute infrastructure since the semiconductor supercycle of the late 1990s. China is sprinting to close the GPU supply gap. In this context, Zhipu's price hike is a capacity-reservation signal, not just a revenue decision. Every chip broker, cloud reseller, and AI industrial-park planner in the country will read 1078 yuan as a claim about the domestic value of reasoning compute. The pricing anchor has cascading effects that run far beyond developer tools.
Regulation adds another layer. Hong Kong's drive to license virtual asset platforms was never purely about innovation; it was about rerouting regional capital flows. Similarly, a high domestic price for AI coding tools is not purely a commercial matter. It feeds the broader narrative of AI self-sufficiency, reshapes procurement expectations in state-linked enterprises, and reinforces the priority of upstream compute investment. The GLM price hike is a micro-signal in a macro-strategy of technological independence.
There is also an unbundling happening in the developer psyche. Chinese engineering culture has long treated domestic tools as a cheaper, imperfect substitute for Western ones. Zhipu just declared that this mental model is obsolete. The message is loud: domestic AI infrastructure is no longer the budget option. Whether the model quality justifies that claim is an open question, but the historical shift in positioning is real. A 1078-yuan Chinese coding seat changes the negotiations in every enterprise procurement meeting from cost-saving to capability comparison.
The consumer-price hike filters for the users who can pay. The mid-August window filters for the users who can commit. The enterprise story filters for the users who can govern. The entire movement is a migration toward a different customer class.
The Ledger Will Answer
The conversion rate between credits and work is the contract nobody has read. Read it before you buy. Demand the credit consumption table. Ask for the cache deletion policy. Benchmark your own month of work. If Zhipu opens the ledger, the next eighteen months could prove this to be a classic refinancing of a scarce asset into a durable enterprise franchise. If the ledger stays opaque, the 538-yuan and 1078-yuan tiers are not price discovery; they are liquidity extraction from a moment of scarcity.
Here is the betting-guidance version. Over the next sixty days, watch three things. First, whether Zhipu publishes typical-task credit examples. That single document determines fairness. Second, whether independent developers in China publish real-world credit-burn rates and cost-per-task comparisons against Copilot, Cursor, and domestic rivals. That data will separate the architectures. Third, whether the V1 window in mid-August is oversubscribed or ignored. That is a direct vote of trust. None of these require a model benchmark. They only require the vendor to speak clearly.
One final provocation. What if Zhipu does not care about retaining the price-sensitive indie developer at all? The entire structure โ scarcity-limited old plans, a credit system that rewards technical sophistication, MCP billing, grandfathered enterprise users โ points toward a purpose-built enterprise AI development platform. The consumer-facing price hike is a wall to keep hobbyists off the cluster. The real product is a governance-ready, code-specialized private workload service. If that thesis is correct, 1078 yuan is not a price point; it is a gate.
Is the price a premium or a tax? Entropy is the only constant in liquid markets. Transparency is the only thing that buys trust in a price shock. The ledger will tell, but only if Zhipu chooses to open it.
