Jev 新闻分诊实测：384 条标题 24.9 秒、19 美分
作者用 Jev 做新闻分诊，384 条标题 24.9 秒出结果、成本 19 美分，对比 Claude Opus 5 只处理 4 条花 77 美分，并解释 Jev 是「只做决策」的架构而非小号 LLM。
The pattern I keep seeing in agent infrastructure right now is that most of what looks like reasoning is actually just classification wearing a costume. TypeSafe AI's Jev makes that split explicit. It doesn't generate text at all. You hand it a state and a typed question, it hands back a calibrated probability over a fixed set of answers. No prose, no chain of thought, just a number. The demo that got me was a news triage task. Jev read 384 headlines in 24.9 seconds and told 15 brands which stories to jump on, for 19 cents. Claude Opus 5, working the same feed, got through 4 of 384 headlines and cost 77 cents. Per headline that's roughly 390x cheaper. TypeSafe AI just raised $40M in seed funding led by DCVC at a $200M valuation. Diogo Almeida spent years at OpenAI and helped invent RLHF, so this isn't someone new to how expensive general models get at scale. My read is this fits a pattern that's older than AI. Redis sits in front of Postgres. A CDN sits in front of the origin server. A router model sits in front of the frontier model. Every layer of computing eventually grows a cheap fast filter in front of the expensive slow thing, and agent orchestration is just now catching up. OpenAI's own GPT-6 Sol and Luna moved the same direction this month, cutting prices 80 to 96 percent against Opus 5 on business workflow benchmarks by trading some ceiling for a lot of floor. Anthropic has Haiku for the same job. The difference with Jev is it isn't a smaller LLM, it's a different architecture built only to decide, route, and score. What agents actually spend most of their calls on isn't writing, it's deciding whether to keep going, which branch to take, whether a result is good enough. Those are classification problems that got routed through million dollar reasoning models because reasoning models were the only tool anyone had. Jev's bet is that once you build the tool for exactly that job, the frontier model stops being the default for tasks that never needed it. https://x.com/elvissun/status/2100951347080421409
这条是引用着另一条写的：「Jev 24.9 秒读完 384 条新闻选题」