OCR + Jev 决策实现每步 0.0002 美元的 computer use
截图经 OCR 转文本，Jev 拆成三个并行 Choice 问题选下一步动作，比 Opus 5 便宜 155 倍、快 14–40 倍。
someone built computer use for $0.0002 per step. no screenshots sent to a big model. ever. here's exactly how it works: screencapture → Vision OCR converts screen to text → Jev reads the text and picks the next action → code executes the click. that's the entire loop. no vision model processing pixels. no frontier LLM writing a plan. just OCR and a typed decision. the cost comparison on the same task: > Jev: $0.0002 per decision / 0.13-0.38s latency > Claude Opus 5: $0.032-0.08 per decision / 5.2s latency 155x cheaper. 14-40x faster. per step. a 12-step task with Opus 5 costs $0.40-0.90. the same task here costs $0.003. the architecture that makes it work: instead of one big "what do I do next" question, the loop asks Jev 3 parallel Choice questions per step: > kind: click, open site, type text, scroll, wait, done? > item: which OCR block to click? > site: which catalog site to open? splitting into 3 questions keeps screen noise out of the action choice. every stall the builder found came from 2 options that meant the same thing confusing the model
这条是引用着另一条写的：x.com/p0lybender/status/2105657133702193458