{"version":"2026-10-19.observations.2","reading":{"id":"reading-cost","title":"榜单费用为什么不是你的完整工作成本？","deck":"把题目/API费用、全部失败尝试、人工核验和最终交付放回同一个范围。","editorialAt":"2026-10-15","sourceIds":["C-AIME","C-LCB","C-EPOCH","C-AA"],"sections":[{"title":"先核对测试的产物","paragraphs":["LCB看历史编程题的执行评分；AIME看固定数学赛事答案；FrontierCode组件是标准化rubric分。它们都提供工具选择线索，但不是同一种业务交付，也不提供相同质量下的完整人时。"],"sourceIds":["C-LCB","C-AIME","C-EPOCH"]},{"title":"一笔费用包含了什么","paragraphs":["AIME费用/题来自同模型、档位、赛事的官方页，仍缺单行实际重复次数、运行日与完整预算。AA Cost per Task方法按测试权重计算输入、缓存、输出token成本，不能转成你的完整工作费用。AA现行条款限制再发布，这里保留方法与原文入口，没有复制数字排行榜。"],"sourceIds":["C-AIME","C-AA"]},{"title":"给自己的试点记一份完整账","paragraphs":["分开记录模型/工具费用、全部失败尝试、人的准备/提示/监督/复核/返工/接管与维护。机器运行、等待和总历时另记，重叠活动不重复累计。人的等待确实占用本人时说明计量约定；不知道就写未知。这个方法建议不是外部论文已替你的团队测出的ROI。"],"sourceIds":[]},{"title":"算术示例不等于已节省","paragraphs":["人工构造例：每次调用0.02美元、10次尝试，调用费为0.20美元；如果仍需人审查30分钟，完整成本还取决于你的人工成本约定与质量验收。示例未执行AI任务，没有代入真实工资或节省比例。只有固定工作量与质量基线后，才比较接收产物的净结果。"],"sourceIds":[]}],"actions":["按同工作量同质量比较","记录失败与人工核验，别只记最终成功调用","把未知费用/时间留为空，不自动推荐最优模型"],"seriesIds":["aime-cost","lcb-code","epoch-code"],"taskIds":["C03","F01","M01"],"frontierIds":["F08","F20","F07","F24"],"evidenceStatus":"active","actualCoverage":null},"sources":[{"id":"C-AIME","title":"MathArena AIME2025固定赛事","url":"https://matharena.ai/","version":"2026-10-14官方模型详情页AIME2025行的有限摘要","locator":"逐模型详情页AIME2025 Accuracy/CI/Cost/Output Tokens","scope":"公开2025题组；模型/推理档位逐行保留","license":"网页数据许可未单独确认；仅少量公布事实摘要与原文链接，不镜像题目/产物","limit":"单行实际运行日与重复次数未知；赛事后模型可能见过公开题；不是研究级证明","publishedAt":null,"readAt":"2026-10-15","collectedAt":"2026-10-14","status":"active","nextReview":"2027-01-13","methods":["https://github.com/eth-sri/matharena"],"family":"C-AIME"},{"id":"C-LCB","title":"LiveCodeBench历史代码生成窗口","url":"https://livecodebench.github.io/leaderboard.html","version":"2026-10-14取得的同一公开JSON；仅overall","locator":"performances_generation.json；默认筛选边界2024-08-01至2025-05-01","scope":"454题；实际题目日期2024-08-04至2025-04-07；execution-based pass@1","license":"代码MIT；成绩JSON数据许可未单独确认；仅有限事实摘要","limit":"不是2026新题实测；模型发布日未可靠建立，全部保持null；不与AIME合并排名","publishedAt":null,"readAt":"2026-10-15","collectedAt":"2026-10-14","status":"active","nextReview":"2027-01-13","methods":["https://github.com/LiveCodeBench/LiveCodeBench"],"family":"C-LCB"},{"id":"C-EPOCH","title":"Epoch ECI固定快照","url":"https://epoch.ai/data/eci-documentation","version":"2026-10-14取得的CSV；ECI与FrontierCode独立指标","locator":"eci_scores.csv；processed_data_for_eci.csv指定benchmark行","scope":"跨测试拟合的综合指数；组件保留模型版本，可能取同名同日不同配置的最高成绩","license":"Epoch数据CC BY 4.0（当前官方链接核对），署名Epoch AI","limit":"模型日不等于测量日；本次快照不是逐日历史；组件0–1分不称工作通过率","publishedAt":null,"readAt":"2026-10-15","collectedAt":"2026-10-14","status":"active","nextReview":"2027-01-13","methods":["https://epoch.ai/data/eci-documentation/methodology"],"family":"C-EPOCH"},{"id":"C-AA","title":"Artificial Analysis成本方法","url":"https://artificialanalysis.ai/methodology","version":"2026-10-15读取的方法与现行条款","locator":"Cost per Task与Intelligence Index；Terms of Use§2.1/2.2","scope":"测试场景的输入/缓存/输出token成本与任务权重；不是完整业务流程费用","license":"现行条款限制复制/再发布，数字排行暂不接入公开投影","limit":"本阶段仅说明指标并链接官方方法；未导入或提供AA数字排行榜","publishedAt":null,"readAt":"2026-10-15","collectedAt":"2026-10-14","status":"active","nextReview":"2027-01-13","methods":["https://artificialanalysis.ai/terms-of-use"],"family":"C-AA"}],"actualCoverage":null}