Tests conducted by Finance Magazine show: For routine office tasks like financial report analysis, the cost of DeepSeek-V4-Flash (official version) is less than 3% of Claude Opus 5. However, in terms of speed, overseas models are noticeably faster. Claude Opus 5 takes an average of 3 minutes 36 seconds, GPT-5.6 Sol averages 4 minutes 33 seconds, and GPT-5.5 averages 5 minutes 49 seconds. Kimi K3 averages 10 minutes 29 seconds, Qwen 3.8-Max (official version) 18 minutes 46 seconds, DeepSeek-V4-Flash (official version) 20 minutes 44 seconds, and Qwen 3.8-Max (preview version) 25 minutes 45 seconds. The specific task was: reading 24 Google financial reports from 2020 to 2025, calculating 10 financial metrics across 20 quarters from 2021 to 2025 (a total of 200 data points), and generating a table. Each of the 18 models underwent two rounds of testing, with each model independently performing the task three times per round, completing a total of 108 tasks. All tests were executed sequentially. Full article: https://mp.weixin.(qq.com)/s/8nmE9mNimSS46JpEHn6ibA