SemiAnalysis 解读 TypeSafe AI 新模型 Jev 的定价与效率逻辑
SemiAnalysis 发文介绍 TypeSafe AI 新模型 Jev，它以 $0.03/M 输入 token、输出免费的价格和效率速度受到关注。Jev 与 LLM 的不同在于返回带校准概率的类型化答案，同时通过在预训练模型上用 Reinforcement Learning for Calibrated Decisions (RLCD) 后训练保留了 LLM 的世界知识。
Jev is a new model from TypeSafe AI which has made headlines for its price and efficiency/speed. It's different from an LLM in that it returns a typed answer with a calibrated probability. However, it's able to retain all the world knowledge of an LLM because it starts with a fully pretrained model and is post-trained with a new technique called Reinforcement Learning for Calibrated Decisions (RLCD) to output probabilities.
Many users are posting about how they've cut down their AI spend by switching over to Jev from frontier models. While cost and time savings are real, they also don't mean what some people think. This doesn't actually change frontier lab spending as agentic workflows still require long-horizon reasoning. At $0.03/M input tokens with free output, they're essentially giving away free otuput tokens because there's no decode loop. The savings measure what it costs to stop generating, not a leap in capability because Jev is only displacing the frontier model that was being used as a router. (3/3)
来源：SemiAnalysis_ · x.com