Chinese AI Model Uses Less Muscle for Coding Tasks

Summary

Z.ai's open-weight GLM 5.2 model (753 billion parameters, 40 billion active) matches or nearly rivals Anthropic's Opus 4.8 in various agentic coding benchmarks, despite having an API price that is roughly one-fifth of Opus 4.8's. Thanks to the MIT license, any organization can host the model on their own infrastructure, avoiding data sovereignty concerns. However, Z.ai's own report acknowledges that GLM 5.2 lags behind Opus 4.8 by at least 10% in more challenging coding benchmarks like SWE-Marathon, where the model completed only 13% of tasks.

Why it matters

The model's near-frontier coding performance, aggressive price advantage, and open weights directly target the two main pillars that support US labs' API revenues: capability and cost. For engineering teams routing routine tasks to cheaper models, GLM 5.2 is becoming a reliable default option. This narrows the addressable market share for premium frontier APIs. The model also accelerates the rate at which Chinese labs are closing the gap with their US counterparts; according to the Stanford AI Index, this closure rate has risen from one-fifth in 2020 to approximately half today.

Highlights

• $4.40 per million output tokens; approximately $22 for Opus 4.8 and $44 for Fable — a 5 to 10x cost difference for the same task class
• MIT-licensed open weights enable self-hosting, eliminating data sovereignty objections often raised against Chinese-origin models
• Z.ai's own benchmark report shows GLM 5.2 winning only two easy reasoning tests and losing to Opus 4.8, and frequently to GPT-5.5, across all coding benchmarks

Critical take

The article's framing of GLM 5.2 as a 'gap-closer' is softened by its own data. In SWE-Marathon, the most challenging long-horizon coding benchmark, the model scored 13% compared to Opus 4.8's approximately 26%. Z.ai's report does not mention Anthropic's newer models, Fable and Mythos. The 'cost-performance revolution' narrative relies heavily on routine task scenarios, where a 5% price drop matters most. However, the frontier-level reasoning gap that justifies premium pricing remains largely intact.