V4.1-Flash 对比 Opus 5：DeepSWE 77.0 对 74.0，Automation-Bench 60
汇总官方数据：DeepSWE 77.0 胜 Opus 5 的 74.0、Automation-Bench 60.5 对 50.3、Agents' Last Exam 35.4 对 28.6，但 HLE、Terminal-Bench、NL2Repo 仍落后；搜索 API 每千次 0.75 美元。
DEEPSEEK UPDATED V4.1-FLASH AND OPENED SEARCH AND FETCH API TO EVERYONE the model got stronger than the last one, solves tasks faster and gives shorter answers, which makes it cheaper to run. the weights are up on Hugging Face, and the new version is live on the website, in the app and via API where it beats Claude Opus 5: - DeepSWE v1.1: 77.0 vs 74.0, with GPT-5.6 Sol at 73.0 - Automation-Bench: 60.5 vs 50.3 - Agents' Last Exam: 35.4 vs 28.6 the jump over the previous Flash is big: > Terminal-Bench 4.0: from 31.2 to 42.3 > ProgramBench: from 20.3 to 29.1 on HLE (56.3), Terminal-Bench 4.0 (51.8) and NL2Repo-Bench (75.3), Opus 5 is still ahead. the numbers come from the company itself, and each benchmark uses its own evaluation method search runs on DeepSeek's own engine: $0.75 / 1,000 calls per endpoint