拆解两篇爆帖：Opus 5.5 与 Astra 组合是否被夸大
作者核对两篇高浏览量帖与对应论文，指出「Opus 5.5 规划 + GPT-6 Astra 探索」的分工是帖子自行拼凑，论文实际单模型运行，3.2 倍省钱等数据在原文中找不到。
Stanford just retired hand-built search scaffolding with Opus 5.5 plus GPT-6.1 Sol, says a 270,088-view post. The paper it matches, from Texas A&M, runs a single agent. A second post, 231,015 views, pulls the same move on a real arXiv ID. I read both posts and both papers. The first sells a code-search loop with no human-designed scaffolding. > DIRECTOR plans each round from the idea graph; the post seats Opus 5.5 here. > PROPOSER writes code variations in sandboxes; the post hands this seat to GPT-6.1 Sol. > IDEA GRAPH stores lineage in Neo4j, so MAP-Elites and MCTS become single Cypher queries. > EVALUATOR never mounts in the agent's container and scores a copy offline. > META-AGENT reads transcripts between search blocks and writes skill files for later branches. In the second post, GPT-6 Astra explores math at T=0.8 while Opus 5.5 audits steps at T=0.0. Where it breaks > claimed: 3.2x lower spend, 0% reward hacks, first in 7 AtCoder contests, says the first post > claimed: MATH500 29.6% to 58.0% with Opus 5.5 auditing Astra, says the second > checked: AgentDiscover runs Opus 5, GPT-6 Sol, Opus 4.6 or Haiku 4.5, one model per run > checked: 7 AtCoder firsts are hypothetical placings vs past results; 153.9x kernel speedup over naive > checked: authors found and fixed reward hacks in inherited evaluators; no 0% claim > checked: 2608.05643 tests Qwen, Ministral and LLaMA; 58.0% is Qwen2.5-1.5B vs a 29.6% reward-model best-of-8 baseline > not public: 3.2x and 91.2%, absent from the paper text I read; no GPT-6.1 Sol release page Both methods run on one model; only the posts split vendors. The graph and the walled-off evaluator are the parts with evidence. Before paying for two frontier APIs, check the paper's authors and model table. My terminal below plays the post's orchestration log next to a what-the-paper-says column ↓
这条是引用着另一条写的：x.com/joyingeth/status/2106050928042238314