DeepSeek V4's Price Hike Meets ZhiPU GLM-5.3: A Battle for the Coding Agent Throne in the AI API Arena

Alerts screamed while the rest of the world slept. DeepSeek V4 raised its peak input price from ¥4 to ¥9 per million tokens. Hours later, ZhiPU GLM-5.3 landed at ¥8 input, ¥28 output—underpricing DeepSeek by ¥1 on input, matching on output. The floor didn't just drop; it shifted. This isn't a simple price war. It's a strategic repositioning of two Chinese AI giants fighting for the same slice of the coding agent market—the most token-hungry, revenue-dense segment in the industry. Let's rewind. DeepSeek V4 was the 'value king'—cheap, fast, and open-source. Developers loved it. Then came the hike. Why now? The answer lies in the infrastructure. DeepSeek's cache hit pricing is ¥0.15 per million tokens during peak—that's 1/60th of its standard input price. This isn't a mistake. It's a signal. DeepSeek has optimized its KV-cache system to a level where the marginal cost of serving cached prompts is near zero. The price hike on fresh inputs is a way to force users toward cache-friendly patterns, smoothing GPU utilization. It's a textbook case of off-peak pricing applied to AI inference. But ZhiPU GLM-5.3's arrival is a direct counterpunch. The announced benchmark scores—9 tasks, 7 wins for GLM-5.3—are all agent and coding focused. Think DeepSWE, Terminal Bench, HLE with Tools. The gap is narrow: 2-4 points in most cases, within statistical noise. Yet the narrative is 'GLM-5.3 is stronger.' The trick is that ZhiPU cherry-picked exactly the battlefields where it leads. No general reasoning, no math, no multilingual. Why? Because DeepSeek V4 likely still dominates those. This is a PR war disguised as a tech war. Now, the real meat: pricing. After the hike, DeepSeek V4-Pro costs ¥9 input / ¥27 output at peak. GLM-5.3 is ¥8 / ¥28. The difference is ¥1 on input—far less than the switching cost for a developer. That means the decision is no longer about price. It's about model capability, ecosystem lock-in, and infrastructure efficiency. DeepSeek's cache pricing (¥0.15) is 13x cheaper than ZhiPU's (¥2). For high-reuse patterns like code completion or template-based agents, DeepSeek is the only rational choice. ZhiPU's cache is a placeholder, not a weapon. But here's the contrarian angle—the hidden story. DeepSeek's price hike might not be a voluntary move. The inference pressure on its MoE architecture could be hitting capacity limits. By raising prices, it's smoothing demand crests. Meanwhile, ZhiPU's GLM-5.3 launch timing is aggressive—riding the wave of user dissatisfaction. Yet, DeepSeek's open-source community is a fortress. Developers can self-host, bypassing API pricing entirely. ZhiPU cannot match that. The real battle is not on the API pricing spreadsheet but on the GitHub repo commit history. In crypto, the news is the asset until it isn't. Here, the asset is the developer mindshare. DeepSeek has the cache infrastructure, the open-source goodwill, and the off-peak pricing flexibility. ZhiPU has a marginally better benchmark score on a curated set of agent tasks, a ¥1 price advantage, and a well-timed PR campaign. The winner will be the one that converts these technical advantages into sticky user habits. Takeaway: Watch the third-party benchmarks. SuperCLUE, OpenCompass, LMArena—real data will surface within weeks. Monitor Cursor, Windsurf, and other coding agent platforms for supplier switches. The tide turns faster than a flash crash. Be ready to move.