The GLM Coding Plan Price Shock: A Tokenomics Autopsy of China's AI Compute Market

I don't normally write about AI coding subscriptions. That's a different beat, a different tribe, a different kind of FOMO than what I live on. But when I saw the numbers — a 261% price hike on a single plan tier, a points system that reads like a gas fee schedule, and a limited-quantity purchase window that smells suspiciously like a token sale — I had to dig in.

Here's the thing nobody in the blockchain space is saying out loud: Zhipu AI just ran a textbook tokenomics experiment on a programming product, and the results tell us more about the future of compute markets than any L1 whitepaper I've read this year.

The transaction trace: Zhipu's GLM Coding Plan is switching from prompt-count limits to a points-based billing system. New user monthly fees jump from 49/149/469 yuan to 118/538/1078 yuan for the Lite, Pro, and Max tiers respectively. That's a 141% increase on Lite, 261% on Pro, 130% on Max. And the new points system converts four resource types — input tokens, output tokens, cache tokens, and MCP calls — into a single metered currency.

The 2017 break didn't teach me to be cynical about protocols. It taught me to read the transaction trace before reading the press release. So let's trace this one.

This is not a story about AI. It's a story about how scarce compute gets priced when demand outstrips supply — and what happens to the people who get caught in the repricing.

The Context: From Gentle Queues to Fee Markets

Let me set the stage for anyone who hasn't been watching the Chinese AI coding tool scene.

Zhipu AI is one of China's top-tier large-model labs. Think of it as the DeepSeek-era darling that actually ships commercial products instead of just publishing papers. Its GLM model family spans everything from chat interfaces to coding assistants, and the GLM Coding Plan was designed to be the subscription wrapper around its code-generation capabilities.

The old plan had a fascinating constraint: because demand was so intense, Zhipu only released a limited number of subscription slots each day, at 10 a.m. sharp. It was a queue. A rationing mechanism. You wanted in, you showed up at the right time and hoped the allocation didn't sell out before you clicked.

That behavior — daily allocation, limited supply, desperate users — is something I've seen before. It's the same energy as an NFT whitelist. It's the same psychological dance as a token sale with a capped raise. Scarcity creates urgency. Urgency creates demand. Demand creates FOMO. And FOMO makes people stop reading the fine print.

Now Zhipu is changing the game. As of the new pricing structure, GLM Coding Plan is moving away from "prompt count limits" — which measured usage in conversation rounds or request counts — toward a unified points system. Every resource you consume, from input tokens to output tokens to cached context to MCP tool calls, gets converted into points. Run out of points, and presumably you pay more or wait for the next cycle.

On the surface, this looks like a simple pricing adjustment. A vendor raising prices because it can. But the structure tells a much more interesting story.

The 2025 EU MiCA regulatory hearings taught me to read intent behind dry legal text. When a company changes its metering granularity at the same time it changes prices, it's not just adjusting revenue — it's redesigning the relationship between the user and the resource.

Core Analysis: Seven Dimensions, One Signal

I'm going to do this the way I was trained to do it as a quantitative analyst during the 2017 Parity multisig crisis: break the event into its components, assess each one, and then ask what the hidden geometry tells us.

Dimension One: The Technical Route — Points as Gas Fees

Let's talk about what a points system actually is, mechanically.

The old GLM Coding Plan limited users by prompt count. That's a crude meter. It measures requests, not resources. One request might be a one-line autocomplete; another might be a full-file refactor with 8,000 tokens of context. Treating them the same was like charging the same gas fee for a simple ETH transfer and a complex DeFi liquidation — the pricing bears no relationship to the actual computational load.

The new system fixes that asymmetry. Input tokens, output tokens, cache tokens, and MCP calls each convert into points at presumably different rates. What does this tell us technically?

First, it tells us Zhipu's backend can now distinguish between different resource classes. That's a nontrivial engineering capability. To meter four different resource types with granular precision, you need request-level accounting that tracks every token in, every token out, every cache hit, and every tool invocation. This is the kind of infrastructure that only gets built when a company is serious about treating compute as a metered utility.

Second, the design reflects a cost model. Output tokens are almost certainly more expensive in points than input tokens, because autoregressive generation — token-by-token prediction — is computationally heavier than processing existing input. The fact that Zhipu separates cache tokens at all tells me the company knows that context caching is one of the biggest levers for inference cost reduction. Encouraging cache reuse by making cached tokens cheap in the points system is a load-management strategy disguised as a user benefit.

Third, and this is the part I find most interesting, MCP calls are explicitly billed. MCP — the Model Context Protocol — is the standard that lets AI models connect to external tools: databases, APIs, code repositories, or other services. By billing MCP calls, Zhipu is signaling that GLM Coding Plan is no longer just a chat-based code generator. It's becoming an agent execution layer. The product is evolving from "ask the model to write code" to "give the model tools to write, test, and deploy code."

I know a thing or two about watching products evolve from simple to compound. Back in the 2020 DeFi summer, I built a Python script to monitor Uniswap V2 reserve changes in real time. The first version just watched token pair ratios; the second version predicted liquidity shifts and produced trade signals. When a product adds metering for tool calls, it's telling you the next layer of the stack is coming.

What's not clear — and this is critical — is the actual conversion rate. How many points does a typical coding task consume? How does a 1,000-token response compare to a single MCP call that queries a GitHub repository? Without a conversion table, users are buying a pig in a poke. It's like joining a DeFi protocol that tells you there are swap fees but doesn't disclose the fee structure until after you've committed liquidity.

The one thing the limited information tells me with high confidence: Zhipu is pulling a classic "gas fee" move. The points system is a variable-rate fee market. When the network is congested, prices rise. When demand is high, users pay more. This is Ethereum's EIP-1559 mechanism, transplanted into a centralized subscription product.

Dimension Two: Commercialization — The Dual-Track Migration

Here's where the story gets genuinely interesting for anyone who understands crypto-native growth tactics.

The new pricing structure doesn't just raise prices for everyone. It grandfathers existing V2 users at their old rates — both individual and team plans — and opens a last-chance purchase window for V1 users to grab the old V2 pricing before the new structure goes fully live. The V1 purchase entry is expected around mid-August.

Let me translate that into crypto terms: this is a token migration with a sunset window. Holders of the old token (old subscription) can convert at the pre-migration price. New entrants pay the post-migration price. And the team announces the migration date in advance, creating a final rush of "cheap" acquisitions before the window closes.

Why does this matter? Because the structure is engineered to do three things simultaneously.

First, it limits churn among existing users. The V2 grandfathered pricing means the people who already committed to the product are not punished for early adoption. That protects the user base and prevents a wave of cancellations that would make the Q3 numbers look terrible.

Second, it creates a conversion deadzone. V1 users who have been on the fence — who have the old plan, who are watching it expire — now face a binary choice: buy the old V2 plan at the pre-hike price before the window closes, or sit out and later face the new prices. That's a loss-aversion trigger. Behavioral economics 101. The window closure is a ticking clock, and ticking clocks are the oldest conversion tool in the book.

Third, the dual-track structure separates the user base into two classes — the privileged incumbents and the high-paying newcomers. This is a deliberate differentiation strategy. It tells the market: "We're raising prices, but we're not punishing loyalty." And in the same breath, it tells new users: "The product is worth 261% more than it was six months ago."

From a pure business perspective, this is textbook price discrimination. I've seen it in micro-cap tokens where the team does a stealth revaluation. The people who got in early get the legacy terms; the new people pay the repriced reality.

What's genuinely concerning is what's not disclosed. The announcement does not clarify whether the higher prices come with higher usage limits. There's no comparison of "points per dollar" between the old and new plans. The customer is effectively asked to trust that the repricing is fair — that the new points system, with its opaque conversion rates, doesn't secretly represent a massive devaluation of purchasing power.

This is the same trust deficit that plagues unaudited smart contracts. If you can't verify the rules, you're betting on the operator's goodwill. And goodwill is not a reliable risk management strategy.

Dimension Three: Industry Impact — The Subsidy Exit

Let me zoom out. Zhipu is not just any company making a pricing decision. It's one of the most prominent Chinese AI labs in the market. Its pricing choices are read by the entire domestic industry as a directional signal.

What signal does a 130-261% price hike send?

The first signal: the era of subsidized AI coding tools in China is ending. The old plan's "limited daily allocation" was a loss-leader mechanism — giving away high-quality code generation to build brand and collection of usage data, while capping total exposure through scarcity. That's the classic "burn to acquire" model. And it's the same playbook I watched DeFi protocols run in 2020: yield farms paying out their own tokens as rewards to attract liquidity, hoping the users would stick around after the subsidies stopped.

Most of them didn't stick. And that's the risk Zhipu is now facing.

The second signal: price-sensitive users will migrate to alternatives, and the migration will happen faster than Zhipu might expect. In crypto, liquidity moves at the speed of social media. Developers, like yield farmers, are mercenaries. Their switching costs are context — the codebase, the habits, the plugin ecosystem — but those costs are not insurmountable.

Let me look at the competition this price change creates. The new GLM Coding Plan Max tier costs 1,078 yuan per month, or roughly $150. GitHub Copilot's individual plan is around $10 per month. Cursor Pro is about $20 per month. That's a 3.8x price premium over Cursor for the equivalent position in Zhipu's product hierarchy.

Is that justifiable? Possibly. If the GLM model's code generation capability is meaningfully better than the alternatives, and if the Chinese-language code ecosystem support is stronger, and if enterprise customers value data residency — then a premium might be defensible. But here's the rub: Zhipu provides no benchmark evidence. No HumanEval results. No SWE-bench numbers. No comparison table showing why the new price is justified relative to the old plan's efficiency.

For an industry that runs on benchmarks — crypto traders trust code, AI users trust evaluations — that's a red flag. It's like a new L1 launching without a public testnet. Yes, the team has a track record. But confidence is not a substitute for verifiability.

There's a third signal I want to flag, and it's the most important one for the broader ecosystem: Zhipu's willingness to raise prices this aggressively tells us that demand for AI code generation is genuinely supply-constrained. The old daily allocation at 10 a.m. wasn't a marketing stunt. The queues were real. The demand was real. And when demand exceeds supply in a product market, the operator eventually converts the scarcity into pricing power.

In crypto terms, this is the moment when an asset transitions from a meme to a productive asset. The question is whether the underlying value can keep pace with the price.

Dimension Four: Competitive Landscape — The Red Sea Price Band

Let me state the obvious: 538 yuan for the Pro tier puts Zhipu in a price band that crosses into premium international territory. In per-dollar terms, it's above the mainstream Western coding tools. That's a bold positioning, and it deserves a closer look.

Zhipu's competitive story rests on a basket of advantages that are more credible than a single headline metric. The company owns the full GLM model stack — from base models to coding-tuned variants — which means it can bundle model calls, context caching, and MCP tool services into a single unified product. A vendor that controls the model weights and the serving infrastructure has a cost structure advantage that pure front-end tools like Cursor (which likely rely on third-party models) do not.

But there's a strategic bet buried in this pricing: Zhipu is betting that "domestic large model + code capability" creates a combination that's irreplaceable in the Chinese ecosystem. For enterprise customers, the value proposition isn't just code quality. It's data compliance. It's the ability to keep code inside China's regulatory jurisdiction. It's private deployment options that satisfy the Data Security Law and the Personal Information Protection Law. Those are constraints that international tools like GitHub Copilot and Cursor cannot easily meet.

That's a legitimate moat. But — and this is a big but — the moat only holds if the underlying code quality is competitive. Chinese developers are not isolated from global benchmarks. They read SWE-bench results. They compare output quality. They know what good looks like because they often work in bilingual codebases.

Here's the competition matrix as I see it:

The price-sensitive refugees from Zhipu's hike have three obvious destinations. Tongyi Lingma, Alibaba's coding assistant, has enterprise cloud backing and aggressive pricing. CodeGeeX, from Zhipu's academic roots in Tsinghua, offers a cheaper domestic alternative. And the international tools — Copilot and Cursor — offer established quality at price points that are now dramatically cheaper than Zhipu's new tiers.

Will they defect en masse? I don't think so, for one reason: habit. In 2021, I watched the Bored Ape floor price lag behind Twitter influencer mentions by minutes, not hours. I built a social alpha arbitrage guide around that observation. The lesson from that period: community momentum and switching costs matter more than price at the moment of repricing. People who have invested months of context into GLM's interface, prompts, and workflows will not leave overnight. But the new user pipeline will dry up if the price-to-performance ratio doesn't justify the premium.

The most dangerous competitor to Zhipu isn't any of those tools. It's the open-source model. If the GLM model's code capability is roughly matched by open-source alternatives that run locally or via cheap API services, then the premium for the "convenience package" of Zhipu's points-based subscription becomes hard to justify.

Dimension Five: The Ethics & Security Layer — Code Data in the Cache

Let me pivot to the dimension that I think too few industry watchers are taking seriously: the security and privacy implications of the new points system.

The keyword here is cache tokens.

For the points system to work efficiently, Zhipu's backend needs to cache user context — the code snippets, the conversation history, the file excerpts that define a coding session. When the points system makes cached tokens cheaper, it's explicitly incentivizing users to allow their contexts to be reused across requests. That's a design choice about how to reduce inference cost, and it's clever. But it raises a question that nobody in the announcement appears to have answered: where is the cache stored, who can access it, and how is it deleted?

In the security landscape, code is one of the most sensitive data types a company possesses. Source code can contain credentials, business logic, proprietary algorithms, and trade secrets. If a user's context is being cached on Zhipu's servers, that cache becomes a target. A breach of cache infrastructure would be far more damaging than a breach of chat logs, because code is machine-readable and directly exploitable.

The MCP billing feature compounds the risk. MCP means that the model connects to external tools. That implies authorization flows — the model is being granted permission to interact with databases, version-control systems, deployment pipelines. If the MCP permission boundaries are not tightly controlled, there's a real risk of third-party data leakage or even unauthorized operations. Imagine an AI coding agent with access to a production database, and a poorly-scoped prompt accidentally triggers a destructive query.

I've audited enough smart contracts to tell you that this is exactly the kind of attacker surface that keeps security researchers awake at night. Every connection point is an entry point. And the product keeps adding connection points.

Now let me be fair. Zhipu is not alone in this. Every AI coding tool with context features faces the same challenge. OpenAI, Anthropic, and Google all handle code data in ways that have security implications. The difference is that Zhipu's announcement centers its commercialization on the points system — the cache tokens and MCP calls — without addressing the security framework.

Is that an unfair criticism? In a vacuum, yes. Pricing announcements don't usually come with security audits. But in the current environment, where AI supply-chain attacks are becoming a documented threat category, the silence is notable.

Here's what I'd want to know before trusting this product with production code: Can users disable code caching? Does disabling it affect points or pricing? Does the product comply with Chinese data law's requirements for code-based data de-identification? And does Zhipu have an audit mechanism for MCP third-party services?

If those answers are not forthcoming — if the position is "trust us" — then this is a governance red flag. In my crypto career, the projects that refused to be transparent about their mechanics were the ones where the hidden mechanisms eventually hurt users.

Dimension Six: Investment & Valuation — Pricing Power as a Signal

Let me step back and ask what this pricing change signals to a different audience: investors.

Zhipu is one of China's most heavily funded AI startups, and its financing trajectory has been under close scrutiny since the post-ChatGPT AI boom began. A pricing strategy that raises revenue per user by 130-261% is, in the most direct terms, a move to improve unit economics.

Venture capital loves unit economics. Show a fund a subscription product with rising ARPU, disciplined cost recovery, and a clear path to gross margin expansion, and you've earned the right to have a conversation at a higher valuation.

The points system is particularly investor-friendly because it represents a shift from flat-rate pricing to usage-based pricing. Usage-based pricing is a feature that VCs in the AI infrastructure space look for. It aligns revenue with actual compute costs, which is the key to sustainable margins.

Here's the subtle calculation I think Zhipu is making: if the subscription base doesn't shrink disproportionately after the price hike, monthly recurring revenue (MRR) will climb steeply. That surge would land right in time for a Q3 fundraising narrative. The V1 user purchase window in mid-August is strategically timed to capture the last cohort of old-price conversions before month-end, pushing revenue recognition into the quarter.

I'm not saying the price hike is purely a fundraising move. The demand factors are real. But the timing is good — almost too good — and the fact that Zhipu has not announced corresponding improvements in model capability alongside the price change tells me the repricing is primarily about revenue optimization, not product upgrades.

In my experience reading market signals — from the 2017 Parity multisig crisis to the Terra collapse in 2022 — the most dangerous setups are the ones where the seller knows something the buyer doesn't. Here, the asymmetry is the absence of usage-benefit comparison. Zhipu is asking users and investors to take on faith that the new price is fair. That faith may be justified. But faith is not a risk model.

The investment angle I want to emphasize is this: the willingness to raise prices in a supply-constrained market is almost always interpreted as a positive signal. It indicates product-market fit, pricing power, and management discipline. The alternative explanation — that Zhipu needs higher revenue because of external cost pressures or regulatory burdens — is less flattering but equally plausible.

Which story wins determines whether this is a bullish or bearish event for the company's valuation.

Dimension Seven: Infrastructure & Compute — Reading the Crypto for GPU Constraints

And now to the part that I find genuinely fascinating: what this pricing change tells us about the physical state of Zhipu's compute infrastructure.

The old plan's "daily 10 a.m. limited release" was not a marketing choice. It was a rationing mechanism. When an AI company caps subscriptions, it's usually because the inference servers literally cannot handle more concurrent users. The queues were a physical constraint expressing itself through product mechanics.

The new points system continues that story. Why introduce granular metering? Because you need to control which requests consume how much compute. Why price 538 yuan for the Pro tier? Because you want to filter out low-value requests and reserve GPU capacity for high-value users. Why single out cache tokens? Because cached operations dramatically reduce inference costs, and you want to rewrite user behavior to keep serving costs manageable.

The infrastructure inference is inescapable: Zhipu is compute-constrained, and the pricing structure is a load-balancing tool.

I saw the same pattern in Ethereum during the 2020-2021 bull run. When block space got scarce, gas prices rose. When gas prices rose, only high-value transactions went through. The fee market sorted the demand. Zhipu is doing the same thing: using points as gas, volume as block space, and price as the congestion signal.

What this means for the broader AI infrastructure market is significant. If a top-tier Chinese AI lab is running into inference cost constraints tight enough to require a 261% price hike, then the GPU supply situation in China is tighter than the public narrative suggests. The US export controls on advanced AI chips clearly play a role. Every pricing decision Zhipu makes is partly a consequence of its hardware reality.

It also suggests something about the future direction of the product: expect Zhipu's investment to go into serving optimization — better batching, smarter context caching, model quantization, hardware efficiency — rather than into subsidized user growth. The era of cheap AI code generation is over wherever compute is supply-constrained.

The Contrarian Angle: This Is Not a Price Hike. It's a Market Creation Event.

Everyone reading this news seems to focus on the same thing: "Zhipu raised prices, and prices are now expensive." But I think that's the wrong frame entirely.

What Zhipu has actually done is launch a private fee market for inference compute. And in doing so, it has admitted something enormous: AI compute has become scarce enough to be priced like a commodity futures contract.

The old plan's daily allocation was a lottery. The new plan is an auction. Winners get code generation; losers can still buy at higher prices. The points system is the exchange rate; the tiered subscription is the derivative instrument; and the V1 sunset window is the secondary market for legacy access.

Here's the counter-intuitive twist: I think this price increase is — in the medium term — a bullish signal for the AI infrastructure complex.

Why? Because scarcity is the root of all value. If Zhipu were confident in abundant cheap compute, it would keep prices low and scale users. The fact that it's willing to raise prices and potentially shrink its user base tells me the compute constraint is not a temporary blip — it's a structural condition. And in any industry where supply is structurally constrained and demand is structurally rising, the firms that own the scarce resource accumulate the surplus.

There's another angle almost no one is talking about: the V1 purchase window is a participation ritual. It asks the early users to signal their commitment by buying at the old price before the window closes. That's a form of proof-of-belief. The users who convert before mid-August are the loyalists. The users who don't convert are, in a sense, being selected out. This has the feel of a token vesting schedule: a mechanism that sorts the true believers from the churn-prone renters.

Zhipu is effectively saying: "Prove you value this product by committing before the price rises. If you do, we'll reward your conviction with legacy pricing. If you don't, you'll pay the market rate later." That's a conversion funnel designed around identity, not just utility. And identity-based retention is much stickier than price-based retention.

Also worth flagging: no competitor has yet publicly reacted. The market is still absorbing the news. But the next move will come from Alibaba, ByteDance, or Baidu — the players with deep cloud resources who can afford to undercut Zhipu's new pricing. If they seize the spillover demand from price-sensitive users, Zhipu's rate hike could end up feeding the competition.

That's the risk embedded in the contrarian thesis. The price increase is rational from a resource-allocation perspective, but it's also a strategic opening. Every user Zhipu prices out is a user up for grabs.

The Takeaway: Watch the Signals, Not the Headlines

So what does this all mean for you? Let me give you the three signals I'll be watching over the next 90 days.

First, does Zhipu publish a points conversion table? If the company releases clear rules about how many points a typical coding task consumes — compare that to the old prompt limits. If the per-task cost under the new system is roughly equivalent to the old plan's dollar-per-task efficiency, then the price hike is just an inflation adjustment. If it's worse, the hike is a stealth devaluation.

Second, watch the V1 purchase window in mid-August. If the queue shows up again, if the window sells out, then demand is resilient. If the window passes with no queue, then the price hike is starting to bite and the community is not as loyal as Zhipu hopes.

Third, watch the competitors. The spillover from Zhipu's price change will land somewhere. When Alibaba, ByteDance, or Baidu launch their promotional responses, we'll know how much market share Zhipu has exposed.

The deeper lesson from this entire event — and I want to leave you with this — is that the AI industry has now crossed the same line that crypto crossed in 2017 and again in 2021: the transition from subsidized adoption to market-clearing prices. When that transition happens, the winners are the operators who own scarce resources, and the losers are the users who mistook subsidies for fundamentals.

I don't know yet which category Zhipu's users will fall into. But I know which category the careful readers of this article should be in: the ones who read the transaction trace before deciding whether to hold, fold, or rotate into the next signal.

The 2017 break didn't teach me to be cynical. It taught me that the speed of your interpretation determines the quality of your position. The GLM points system is live. The market is repricing. Move accordingly.