Databricks 拋棄 Claude 轉向 GLM —— 以低 34% 的成本匹敵 Opus
databricks 將 glm 5.2 設為預設編碼模型，因其在其數百萬行程式碼庫的內部基準測試中與 anthropic 的 opus 4.8 效能相當，但單次任務成本僅為 1.28 美元（對比 1.94 美元），這表明中國開源模型正在生產環境中取代西方專有模型。
databricks switched its default coding engine to glm 5.2, finding it statistically tied with anthropic's opus 4.8 at $1.28 per task compared to opus's $1.94, and now routes work by complexity to optimize the pareto frontier of price and performance. internal tests showed top-tier models from openai, anthropic, and open source achieving 82-90% pass rates, but extreme cost disparities—often 60-90% lower for chinese models—are driving western enterprises like databricks, coinbase, and lindy to defect from proprietary incumbents. this validates that open-source models are production-grade and suggests future engineering stacks will dynamically route tasks across a fragmented vendor landscape based on real-time cost-to-quality metrics rather than a single flagship provider.