The ledger shows a discrepancy of 13x. DeepSeek V4’s cache-hit price sits at ¥0.15 per million tokens. Zhiyu GLM-5.3 charges ¥2 for the same cache service. That gap is not a rounding error. It is a structural signal. In the past seven days, the API pricing duopoly of China’s AI sector has shifted from a narrative of “cost killer vs. premium performer” to a deadlock where margins are now measured in single yuan. But the real story is not the price tags — it is the infrastructure asymmetry hidden beneath them.
Context: The Two Titans and Their Price Reset
DeepSeek, the open-weight darling of the Chinese AI developer community, raised its peak-hour API prices for V4-Pro by a significant margin on May 14, 2026. Input: ¥9/million tokens. Output: ¥27/million tokens. The same day — or within hours — Zhiyu, the state-backed AI unicorn, rolled out GLM-5.3 with a near-identical pricing: ¥8 input, ¥28 output. The difference is exactly ¥1 in each direction. This is not a coincidence. It is a textbook value-based pricing attack, executed with surgical timing.
Both models target the same high-margin use case: coding agents. These are the AI-driven tools that autonomously write, debug, and deploy software — the most token-intensive application in the current market. A single agent session can burn through millions of tokens. At peak prices, a typical 5.5M token task costs ¥58.5 on DeepSeek V4-Pro and ¥54 on GLM-5.3. The ¥4.5 gap is negligible for enterprises but meaningful for small teams. And more importantly, it eliminates the price advantage that once made DeepSeek the default choice for cost-sensitive developers.
But the real battle is not fought on mainline pricing. It is fought on the fringes: off-peak discounts and cache-hit pricing. DeepSeek offers an off-peak half-price mode (¥4.5 input, ¥13.5 output) and a cache-hit rate of ¥0.15 during peak hours (¥0.3 off-peak). Zhiyu’s cache-hit pricing is ¥2 — 13 times higher, and its off-peak pricing is not even mentioned in the leaked data. This is where the audit gap widens.
Core: The Systematic Tear-Down — Selective Benchmarks, Asymmetric Infrastructure
1. The Price Pairing Is a Trap
At first glance, the ¥1 difference in mainline prices suggests that the decision factor for developers has shifted entirely to model capability. But my audit of the pricing structure reveals a deeper calculation. The ¥1 gap is well below the switching cost for any serious developer: migration of code, adaptation of toolchains, re-testing of agent workflows. Therefore, the price war is not a war at all. It is a signal that both companies understand that the lock-in effect is already in place. They are not competing on price; they are competing on the sustainability of their infrastructure.
2. The Cache Price Disparity: A 13x Infrastructure Gap
DeepSeek’s cache-hit price of ¥0.15 is 1/60th of its peak input price (¥9). Zhiyu’s cache-hit price of ¥2 is 1/4th of its peak input price (¥8). This ratio is a direct measure of KV-cache system efficiency. A 1/60 ratio implies that DeepSeek’s infrastructure can serve a cached token for less than 2% of the cost of a full computation. This is only possible if the platform has achieved extremely high prefix reuse rates, likely through an optimized attention cache architecture and a scalable retrieval system. Zhiyu’s 1/4 ratio suggests a more conventional caching layer — perhaps a simple LRU cache with limited reuse.
Yield trap detected. Developers who optimize their applications for DeepSeek’s cache system will find it nearly impossible to switch to Zhiyu without a massive cost penalty. The 13x difference in cache pricing is a moat that Zhiyu cannot cross in the short term.
3. The Selective Benchmark Problem
Zhiyu published a comparison table showing GLM-5.3 winning 7 out of 9 benchmarks, all in the agent/coding domain. The wins are narrow: 66.9 vs 62.7 on DeepSWE, 28.5 vs 25.7 on HLE with Tools, 62.5 vs 60.0 on CyberGym. The margins are within the statistical noise of most LLM evaluations. But the real issue is the selection. The 9 benchmarks include zero general language understanding, zero math reasoning, zero multilingual capability. This is a deliberate framing.
Audit gap confirmed. The omission of benchmarks like MMLU, GSM8K, or HumanEval (the classic coding benchmark) is a red flag. If GLM-5.3 were genuinely superior across the board, Zhiyu would have published a full comparison. The fact that they chose only agent benchmarks suggests that DeepSeek V4 may still lead in other areas — or at least be competitive enough to avoid embarrassment.
4. The Off-Peak Pricing as a Capacity Signal
DeepSeek’s off-peak half-price strategy is not just a discount. It is a capacity management tool. By pricing down to ¥4.5 input during low-demand hours, DeepSeek is explicitly admitting that its GPU cluster is underutilized during certain periods. This is a classic peak-load pricing mechanism, common in utilities but rare in AI API services. It implies that DeepSeek’s inference load is highly concentrated in business hours, likely because its user base is dominated by professional developers in similar time zones. Zhiyu, by not offering a comparable off-peak discount, may either have a more balanced load profile or a less sophisticated scheduling system.
Mathematical collapse verified. If DeepSeek’s peak load continues to grow, the off-peak discount may not be sufficient to flatten the curve. The company may need to raise peak prices further or invest in more GPU capacity. The current ¥9 price point may be a pressure test, not a final equilibrium.
Contrarian: What the Bulls Got Right — Hidden Strengths on Both Sides
DeepSeek’s infrastructure moat is real. The 13x cache price advantage is not a marketing gimmick. It represents years of engineering investment in attention reuse and efficient retrieval. For high-repetition patterns — code completion, template-based agent tasks, API call chaining — DeepSeek will remain the cheaper option even after the mainline price increase. This is a classic “razor and blades” strategy: give away the razor (cache) cheap, sell the blades (peak tokens) at a premium.
Zhiyu’s timing is aggressive but defensible. Attacking immediately after a price increase is a classic PR move. But more importantly, Zhiyu’s strength in enterprise and government contracts — where pricing is often negotiated privately — means that the public API price war may be a distraction. The real revenue for Zhiyu comes from custom deployments, not the public API. The ¥8/¥28 pricing is a beachhead to capture the developer market, not a core profit center.
The benchmark bias may be intentional, but not dishonest. Zhiyu is focusing on the fastest-growing segment: coding agents. If an enterprise cares about agent performance, the 7-out-of-9 win is a valid signal. The missing benchmarks are relevant only if the developer needs general-purpose capabilities. For a pure coding agent provider, the selection is justifiable. The issue is that Zhiyu presents it as a holistic comparison when it is not.
The real winner may be the neutral layer. Both DeepSeek and Zhiyu are forcing developers to choose sides. This creates a market opportunity for middleware that routes requests to the cheapest or best-performing model in real time. Think of it as a decentralized exchange for AI compute — a model router that optimizes for cost, latency, and quality simultaneously. The infrastructure players (like Together AI, Fireworks, or a Chinese equivalent) could emerge as the true beneficiaries of this price war.
Takeaway: The Next Move Will Be Infrastructure, Not Pricing
DeepSeek raised prices, and Zhiyu matched. The market is now in a state of pseudo-equilibrium where mainline prices are identical, but the real battlefield is hidden in cache and off-peak pricing. Developers who fail to audit their own usage patterns — cache-hit rates, off-peak shift potential, agent task composition — will overpay by up to 13x.
The question is not whether GLM-5.3 is stronger. It is whether Zhiyu can close the infrastructure gap before DeepSeek releases its next model iteration. The clock is ticking. And the ledger does not lie.