Nomura and Daiwa CEOs said Japan’s stock rally should persist through 2027 but cautioned AI-driven investment could reverse the trend. Nomura CEO Kentaro Okuda said market levels, share prices and corporate earnings have been powered by AI investment and that a shift in AI sentiment would be a major downside risk; he forecasts the Nikkei 225 around 75,000 by year-end and above 80,000 by end-2027. Daiwa CEO Akihiko Ogino expressed similar concern and expects the Nikkei 225 to reach 80,000 this ye

A-share banking sector strengthened intraday; Bank of China, ICBC and Bank of Hangzhou hit intraday record highs, while China Everbright Bank, Ping An Bank, Bank of Beijing and Bank of Communications were among the top gainers.

一項最新的社區測試引發關注：同樣使用GPT-6模型，Codex提供的推理預算可能遠低於直接調用OpenAI API，即使選擇相同的推理強度，差距依然顯著。 在最高推理檔位（Max）下，GPT-6 Astra通過API調用時的推理預算達到832，而Codex僅爲128，相差約6.5倍；GPT-6 Sol的差距更加突出，API端達到768，Codex端只有62，相差約12.4倍。社區測試還顯示，這種差異並非僅存在於最高檔位，在其他可比較的推理強度設置下，Codex的推理預算也普遍低於API端。 所謂推理預算，通常指模型內部推理過程可使用的計算預算，並不等同於實際消耗的Token數量。OpenAI官方文檔確認，推理強度會影響模型的思考深度、響應速度和Token消耗，但尚未證實上述832、128、768和62等具體預算數值。 這一差異可能影響複雜軟件工程任務的表現。對於簡單代碼生成和常規修改，較低的推理預算有利於提高響應速度、控制計算成本；但在大型代碼庫重構、跨模塊依賴分析、複雜故障排查以及長鏈條任務規劃中，更充足的推理預算可能幫助模型探索更多方案、檢查潛在錯誤並完成更深入的驗證。