This week Zhipu pushed the discount price of its flagship GLM-5.3-Flash down to 4.5 cents per task — 1/40th of Anthropic's comparable Claude Opus 4.8. The "frontier = expensive" industry default has been rewritten for the first time by a Chinese company, holding all three cards at once: performance, price, and open source.

What this is

GLM-5.3-Flash is Zhipu's open-source model with 320B total parameters and 18B active parameters (i.e., only 18B parameters are activated per inference), and the first natively multimodal model in the GLM-5 series (it handles images and text directly, without external modules). Read the following together:- 57 on the Artificial Analysis composite intelligence index, tying Claude Opus 4.8;- MIT-licensed open source (the most permissive license: free for commercial use, modification, and redistribution), with HuggingFace, API, and Coding Plan launched simultaneously;- 1M token context window (roughly a medium-thickness book of text per session) out of the box;- Before launch, the anonymous model "Niúlái" took the #1 call-volume spot on both OpenCode and OpenRouter, with all traffic running on a domestic-chip cluster.

Industry view

The editorial desk is hearing three voices.Supporters argue the "frontier = expensive" default has been broken. The combination of sparse-plus-linear hybrid attention, IndexPool cache compression, and mHC scaling efficiency means the 1/40 price gap comes from architecture, not subsidies; domestic chips carrying frontier traffic is treated as a compute-validation moment.Cautious voices flag two risks: 4.5 cents is a limited-time discount; whether costs hold after returning to list price is an open question. Zhipu's own team acknowledges that KV cache memory (historical context stored during inference) is still higher than Kimi-K3 and DeepSeek-V4-Flash.The cooler view: performance parity does not equal experience parity; the AA index is a composite score. On pure-vision benchmarks like BabyVision, there is still a gap behind GPT-5.6 Terra and Gemini 3.7 Flash — the strength is in tool-using coding, not raw visual perception.

Impact on regular people

For enterprise IT: API cost structures may reshuffle. Frontier models that used to strain monthly budgets can now sustain all-night AI Agent runs (programs that execute multi-step tasks autonomously) or bulk long-document processing.For working professionals: the marginal cost (the extra cost of handling one more document) of intensive workloads like full long-document review and report drafts drops sharply; the personal ceiling is moved back by price.For consumer markets: the price pressure from open source combined with domestic chips will transmit. Overseas large models' pricing in Chinese-language markets is very likely to adjust over the next few months.