用 Jev 做分流层，8 小时活压到 48 分钟
多模型流水线实测：Jev 只输出选择/评分/是否，比 Opus 5 便宜 22 倍，不确定的 1/4 才转给 Opus。
OPUS 5.5 + SONNET 5.5 + GPT-6 ASTRA + JEV + DOTS on one spider, and my 8-hour day took 48 minutes... the run is a demo. the prices are real. the cheapest agent on that web is Jev. it only answers with a choice, a score, or a yes/no. 22 times cheaper than Anthropic's Opus 5, 13 times faster, only 3 points less accurate. it cannot write a single sentence. the startup is TypeSafe. the CEO, Diogo Almeida, co-wrote InstructGPT at OpenAI, the paper that taught GPT to follow instructions. now he builds a model that never talks. OpenRouter tested it on 3,080 banking support questions. Jev: 81.0% right, 175 ms. Opus 5: 84.4% right, 2,266 ms. full run: $0.34 vs $7.44. Opus 5.5 was not in the test. then the smart part. Jev answers when it is sure. about 1 in 4 goes to Opus. result: 84.0% accurate for $0.69 per 1,000 questions instead of $2.42. that is my spider. OpenAI's Dots reads 4,860 items. Jev decides what matters. 1 hard case goes to Opus 5.5. Sonnet 5.5 writes 14 files. GPT-6 Astra ships it. most of your day is not work. it is sorting. most people who say they have no time are reading 4,860 things that never needed them.
这条是引用着另一条写的：x.com/noisyb0y1/status/2105964235087856080