三个情报台的每日动态，只收已经核过的。每条注明日期、出处与证据等级，标题点开就是一手来源。没核实的，不上这页。Daily signals from the three intelligence desks, verified before they land here. Every item carries its date, source and evidence grade, and every headline links straight to the primary source.
每日更新 · 最近更新 Updated daily · Last updated 2026-11-07
这里是新闻：三个情报台的每日动态流水，注明日期、出处与证据等级，核过才上。要看每周汇编，去周刊 →；要读单篇完整研究，去深研 →。This is News: a daily stream from the three desks, each item dated, sourced and graded, posted only after verification. For the weekly digest, see the Weekly →; for a single study in full, see Research →.
订阅新闻Subscribe to News
三个情报台有新动态时，发到你的邮箱。不等周刊，新闻单独寄。Get an email when any of the three desks posts something new. News only, separate from the Weekly.
邮箱只用于寄送新闻通知，随时可退订。隐私政策Your email is used only for news alerts. Unsubscribe any time. Privacy Policy
已收到，有新动态会发到你的邮箱。Done. New items will land in your inbox.
刚才没发出去，请稍后再试，或到纠错与联系页订阅。That did not go through. Please retry, or subscribe on the Contact page.
SpaceXSI（原 xAI）开发者文档《Output Watermarking》（2026 年 10 月 31 日更新）写明：自 2026 年 11 月 7 日起，grok-4.7、grok-4.6、grok-4.5、grok-4.3 与 grok-4.20 的 reasoning、non-reasoning、multi-agent 三个变体共七个模型，经 Responses API 与 Chat Completions API 在全部端点（全球与区域）生成的文本默认带不可见水印，结构化输出与工具调用参数也在内。水印靠调整 token 选择实现，不添加任何字符，开发者可从响应头 x-watermark 是否为 true 判断；厂商称不影响质量与速度，此为厂商口径，未经独立测量。退出只能由团队管理员在控制台逐模型操作，约五分钟生效并记入审计日志，且须确认存在法律例外并自行承担其余披露义务。厂商称此举是为遵守欧盟《人工智能法》，与其第 50 条的透明度义务对应（该义务 2026 年 8 月 2 日起适用，存量系统过渡期至 2026 年 12 月 2 日；文档原文只说为合规，未点名条号）。文档另注明，个别模型可能被锁定为不可关闭。边界：文档只覆盖 API，未提及 X 上的 Grok、grok.com、App 与 Grok Imagine，目前也没有公开检测器。本条据多篇逐项引述该文档的报道交叉核对，文档原页未逐页亲读，等级列多源已核。SpaceXSI (formerly xAI) developer documentation, Output Watermarking (last updated October 31, 2026), states that from November 7, 2026, text generated by seven models, grok-4.7, grok-4.6, grok-4.5, grok-4.3 and the grok-4.20 reasoning, non-reasoning and multi-agent variants, carries an invisible watermark by default across the Responses API and Chat Completions API on all endpoints, global and regional, including structured outputs and tool-call arguments. The watermark shapes token selection and adds no characters; developers can check the x-watermark response header for a value of true. The vendor says quality and speed are unchanged, a vendor claim not independently measured. Only team admins can opt out, per model, in the Console; changes take effect within about five minutes, are audit-logged, and require confirming that a legal exception applies and accepting sole responsibility for remaining disclosure duties. The vendor gives EU AI Act compliance as the reason, corresponding to the Article 50 transparency duty, applicable since August 2, 2026, with a transition for pre-existing systems running to December 2, 2026; the docs themselves do not name the article. The docs also note that some models may be locked on. Scope limit: the docs cover the API only and say nothing about Grok on X, grok.com, the apps or Grok Imagine, and no public detector exists. This item is cross-checked across several reports quoting the documentation in detail; the docs page itself was not read page by page, so the grade is multi-source verified.
来源：Source: SpaceXSI (xAI) Developer Docs, Output Watermarking (updated Oct 31, 2026), as reported in detail by Metarix and others
Physical Superintelligence（PSI）2026 年 11 月 6 日宣布：其内部研究系统调用一个未公开的 Anthropic 模型，在运行第三天得到 Aaronson-Ambainis 猜想的完整证明，Lean 证明助手已接受；系统随后两次改进结果，最终界为 Inf_i(f) ≥ c·v^(5/2)/d^6，即 (v/d)^7 量级。该猜想由 Aaronson 与 Ambainis 于 2008 年提出，关乎量子计算在无结构随机输入上能否指数级快于经典计算，证明成立意味着不能。工作由 PSI 技术人员 David Turturean 牵头，他开发并运行该系统、对论文正确性负责，Jim Halverson 审计了证明，Phiala Shanahan 评论了 Lean 形式化。论文与 Lean 4 形式化已在 GitHub 公开（Apache 2.0），据仓库说明，形式化含 648 条经机器核查的定理、无 sorry 占位符、无新增公理。2026 年 8 月证明了该猜想一个特例的斯坦福团队（Blanc、Docter、Strassle、Tan）已读过证明、给出评论，并正撰写更短的专家版说明，称其为重大结果。未决边界照实写：尚无同行评议；供人读的论文本身未经机器核查，关键待核点是 Lean 中的陈述是否与原猜想完全对应，截至发稿 Aaronson 与 Ambainis 本人未公开表态；模型与系统不公开，运行不可复现，可核查的是证明工件而非方法。Physical Superintelligence (PSI) announced on November 6, 2026 that an internal research system running a non-public Anthropic model found a complete proof of the Aaronson-Ambainis conjecture during the third day of its run, and that Lean, the proof assistant, accepted it. The system then improved the result twice, reaching a final bound of Inf_i(f) at least c times v^(5/2) over d^6, on the order of (v/d)^7. The conjecture, posed by Aaronson and Ambainis in 2008, concerns whether quantum computers can be exponentially faster than classical ones on unstructured random inputs; the proof settles that they cannot. The work was led by PSI Member of Technical Staff David Turturean, who built and ran the system and is responsible for the paper's correctness; Jim Halverson audited the proof and Phiala Shanahan commented on the Lean formalization. The paper and the Lean 4 formalization are public on GitHub under Apache 2.0, and the repository description reports 648 machine-checked theorems with no sorry placeholders and no added axioms. The Stanford group (Blanc, Docter, Strassle and Tan) that proved a special case in August 2026 has read the proof, sent comments and is writing a shorter exposition, calling it a huge result. Open limits, stated plainly: there is no peer review yet; the human-readable paper is not itself machine-checked, and the key outstanding check is whether the Lean statement matches the original conjecture exactly; as of publication neither Aaronson nor Ambainis had commented publicly; and the model and system are private, so the run cannot be repeated. The artifact is checkable, the method is not.
来源：Source: Physical Superintelligence (PSI), A proof of the Aaronson-Ambainis conjecture (Nov 6, 2026)
Google DeepMind 科学副总裁 Pushmeet Kohli 2026 年 11 月 6 日以《Advancing the frontier of cryptanalysis》为题公布：由 Dj Dvijotham 与 Abhradeep Thakurta 牵头，与加州大学圣迭戈分校及 Inria 的 Nadia Heninger、Inria 的 Emmanuel Thomé 合作，AI 智能体在人类监督下数周内发现旧式 1024 位 RSA 的两处弱点，并附两篇工作论文。其一为签名伪造：攻击者控制被签数据块约一半以上时，填充几乎不再提供保护；智能体的实现在一把已被解出的 RSA-1024 基准密钥上复现伪造，CPU 时间省 3.5 倍，所用算力不到分解该密钥所需算力的千分之一；对 32 字节指纹一类短输入，新算法约 1600 倍的改进是理论估计，不是实测。其二为批量分解：新算法把每多分解一把 1024 位密钥的平均成本降至约原来的三十分之一。边界由 DeepMind 本人划明：结果只涉及已被淘汰的 1024 位 RSA，2048 位及以上密钥不受影响，未攻击任何在用系统；NIST 自 2013 年起已不接受 1024 位 RSA 用于新签名。本条据多家详细报道交叉核对，DeepMind 原帖与两篇论文尚未逐页亲读，等级列多源已核。Pushmeet Kohli, Google DeepMind's vice president of science, announced on November 6, 2026, under the title Advancing the frontier of cryptanalysis, that AI agents working under human supervision found two weaknesses in legacy 1024-bit RSA within weeks. The work was led by Dj Dvijotham and Abhradeep Thakurta with Nadia Heninger of UC San Diego and Inria and Emmanuel Thomé of Inria, and is described in two working papers. First, signature forgery: once an attacker controls roughly half or more of the signed block, padding adds virtually no protection; the agents' implementation reproduced a forgery on an already-solved RSA-1024 benchmark key with 3.5 times less CPU time and under one-thousandth of the compute needed to factor the key itself; for short inputs such as a 32-byte fingerprint, a new algorithm's roughly 1,600x improvement is a theoretical estimate, not a measured result. Second, batch factoring: a new algorithm cuts the average cost of factoring each additional 1024-bit key to about one-thirtieth. DeepMind itself draws the boundary: the results affect only obsolete 1024-bit RSA, 2048-bit and larger keys are unaffected, no live system was attacked, and NIST stopped accepting 1024-bit RSA for new signatures in 2013. This item is cross-checked across multiple detailed reports; DeepMind's original post and the two papers were not read page by page, so the grade is multi-source verified.
来源：Source: Google DeepMind announcement by Pushmeet Kohli, Advancing the frontier of cryptanalysis (Nov 6, 2026), cross-checked across multiple detailed reports
Anthropic 2026 年 11 月 6 日发布工程文章《The components of agent memory》（Lance Martin 撰写）：把智能体记忆定义为会话之间保留的、来自经验的知识（不含权重更新），分四层，自下而上为会话记录、记忆笔记、记忆表征（如维基）与注入上下文；dreaming 指用离线算力在会话之间整理记忆，可纠错、建表征、把笔记上提、聚焦目标与补缺。在 Continual Learning Bench（CL-Bench）上，仅告知有记忆文件夹可用，Opus 4.7 与 Opus 5 就明显提分；提示 Sonnet 4.6 先读记忆文件再答题并自测其理论后，其成绩接近翻倍，而给强模型定「只记失败」一类固定规则反而削弱了记忆带来的提升。以上均为 Anthropic 自家评测，无外部复现。在约 35 万份内部文档、约 4.2 亿 token 的语料（来自 Anthropic 2026 年 5 月收购的 Stainless）上建的维基，随 dreaming 场次由 2 场增至 128 场，在 25 道问题上的评分覆盖率由 13% 升至 48%；仅搜索为 53%，搜索加维基则随维基长大由 54% 升至 61%。《宝可梦 红》实验中，Fable 5 只留 9 个记忆文件加一个顶层 global.md、得四枚徽章；Sonnet 3.7 在约 10,000 步时写下 413 个记忆文件、分布于 225 个文件夹、只得一枚徽章；一条把宝可梦中心误记在月见山东侧的位置笔记被 dreaming 对照会话记录纠正，纠正后的运行最终到达华蓝市，未纠正的运行三次掉回原路、始终未到。Claude Code 只加载 MEMORY.md 的前 200 行或 25KB。边界：文章是研究与内部实践记述，附有 cookbook 与开源代码，dreaming 未宣布为 Claude 或 Claude Code 的正式功能，Anthropic 自称在记忆系统设计上还有很多要学。Anthropic published the engineering article The components of agent memory on November 6, 2026, written by Lance Martin. It defines agent memory as knowledge gained from experience and retained between sessions, explicitly excluding weight changes, in four layers from bottom to top: transcripts, memory notes, memory representations such as wikis, and injected context. Dreaming is the use of offline compute between sessions to organize memory: correcting it, building representations, promoting notes, focusing on a goal and filling gaps. On Continual Learning Bench (CL-Bench), simply telling Claude a memory folder was available gave Opus 4.7 and Opus 5 a noticeable boost; prompting Sonnet 4.6 to read its memory files before answering and to test its own theories nearly doubled its benchmark score, while fixed rules for strong models, such as record only failures, removed much of the gain. All results are Anthropic's own evaluations, with no outside replication. A wiki built over a corpus of about 350,000 internal documents, roughly 420M tokens, from Stainless, a company Anthropic acquired in May, rose in graded coverage on 25 questions from 13 percent at two dreaming sessions to 48 percent at 128 sessions; search alone scored 53 percent, and search plus the wiki rose from 54 to 61 percent as the wiki grew. In Pokémon Red, Fable 5 kept nine memory files plus a top-level global.md and earned four badges, while Sonnet 3.7, at about 10,000 steps, had written 413 memory files across 225 folders and earned one badge; a note wrongly placing the PokéCenter on the east side of Mt. Moon was caught by dreaming against the transcript, and the corrected run eventually reached Cerulean City, which the uncorrected run never did. Claude Code loads only the first 200 lines or 25 KB of MEMORY.md. Boundary: the article is a research and internal-practice write-up with cookbooks and open-sourced code; dreaming is not announced as a generally available feature of Claude or Claude Code, and Anthropic says it still has a lot to learn about designing memory systems.
来源：Source: Anthropic, The components of agent memory (Nov 6, 2026)
AMD 2026 年 11 月 6 日在 ROCm 博客发布 PrimusClaw 并开源（代码在 GitHub 的 AMD-AGI/PrimusClaw 仓库）：这是一个托管智能体运行时，由会话存储、智能体执行框架与每会话独立沙箱组成，跑在客户自己的 Kubernetes 集群里，贴着 AMD Instinct GPU 部署。AMD 称其自 2026 年 4 月起已在内部生产使用：两套部署累计托管超过 40,000 次智能体会话，服务 70 多名用户，累计逾 100,000 个智能体小时、400 多万次工具调用与 5,000 多亿 prompt token，主部署单次运行中位时长 3.4 小时。代码以 MIT 许可发布，沙箱控制面除外，那部分是 Kubernetes SIG Apps 的 agent-sandbox 项目的 Apache 2.0 分叉；协议上入站支持 A2A v1.0，双向支持 MCP。上述用量数字为 AMD 自报的生产统计，未经第三方核查。AMD published PrimusClaw on its ROCm blog on November 6, 2026 and open-sourced it (code in the AMD-AGI/PrimusClaw GitHub repository). It is a managed agent runtime made of a session store, an agent harness and per-session sandboxes, run in the customer's own Kubernetes cluster next to AMD Instinct GPUs. AMD says it has been in internal production use since April 2026: two deployments have hosted more than 40,000 agent sessions for more than 70 users, totaling over 100,000 agent-hours, more than 4 million tool calls and more than 500 billion prompt tokens, with a median run of 3.4 hours on the main deployment. The code is MIT licensed except the sandbox control plane, a fork of the Apache-2.0 agent-sandbox project from Kubernetes SIG Apps; it supports A2A v1.0 inbound and MCP in both directions. The usage figures are AMD's own production statistics and have not been independently verified.
来源：Source: AMD ROCm Blog, PrimusClaw: An Open-Source Managed Agent Runtime for AMD Instinct GPU Clusters (Nov 6, 2026)
据 Bloomberg 报道并经 Reuters、TechCrunch 等转述：OpenAI 新一轮融资已于 2026 年 11 月 6 日完成首关，约 110 亿美元到账，对应估值 1.25 万亿美元；整轮尚未关闭，未来数周可能增至约 200 亿美元。两名投资人向 Reuters 表示 SoftBank 与 MGX 参与了本轮。背景：9 月底 Bloomberg 曾报道本轮目标为 300 亿美元，现口径已明显下调；OpenAI 财务总监 Sarah Friar 11 月 6 日在 CNBC 称融资仍在进行中。对照锚点：2026 年 3 月一轮为公司正式宣布的 1,220 亿美元、投后估值 8,520 亿美元。本条全部数字来自知情人士转述，OpenAI 未发布公告、无一手文件，等级列媒体转述，照生成器规则不进首页最新动态。Bloomberg reported, with the account carried by Reuters and TechCrunch, that OpenAI's new funding round reached a first close on November 6, 2026, with about $11 billion wired at a $1.25 trillion valuation. The round has not closed and could grow to about $20 billion in the coming weeks. Two investors told Reuters that SoftBank and MGX took part. Background: Bloomberg reported in late September that the round targeted $30 billion, so the current figures are well below that target; OpenAI CFO Sarah Friar said on CNBC on November 6 that the round is ongoing. Reference point: the March 2026 round, formally announced by the company, was $122 billion at an $852 billion post-money valuation. Every figure in this item comes from source-based reporting; OpenAI has published no announcement and there is no primary document, so the grade is media report, and under the generator rule it stays out of the homepage latest four.
来源：Source: Bloomberg reporting, via Reuters and TechCrunch (Nov 6, 2026)
韩国 Lunit 2026 年 11 月 6 日宣布其医疗基础模型项目完成，并把 Learning Unit 2.0（模型号 L1-30B-A5B）的权重放到 Hugging Face 开源（Apache 2.0）：该模型由 23 家企业、高校、研究机构与医院组成的联盟共建，为稀疏 MoE 架构（GravityMoE，源自 Trillion Labs 的 Gravity-30B-A5B-Base），总参数 295.6 亿、每 token 激活 53.4 亿，52 层，上下文 131,072 token，原生面向韩国临床场景训练，设计为在医院内本地部署。模型卡上的基准数字用 Lunit 自研的 CoEval 框架自测。据 Lunit 公布的九家医疗机构半年实测：多学科（肿瘤）诊疗中，86.2% 的 AI 治疗方案被评为临床有效或适当；急诊分诊（KTAS）与药物不良反应因果评估两项，与专家判断相差一级以内的比例为 98%，注意这不是完全一致。监管边界照实写：该模型不是经韩国 MFDS 获批或清关的医疗器械，各机构临床部署前须自行验证，不得写作已获批。本条模型规格以模型卡一手为准，实测数字为 Lunit 公布口径，等级列多源已核。South Korea's Lunit announced on November 6, 2026 that its medical foundation model project is complete and released the weights of Learning Unit 2.0, model ID L1-30B-A5B, on Hugging Face under Apache 2.0. The model was built by a consortium of 23 industry, academic, research and hospital institutions. It is a sparse Mixture-of-Experts model (GravityMoE, post-trained from Trillion Labs' Gravity-30B-A5B-Base) with 29.56 billion total parameters and 5.34 billion active per token, 52 layers and a 131,072-token context, trained natively for the Korean clinical locale and designed to run on-premises inside hospitals. Benchmark figures on the model card come from Lunit's own CoEval framework. In field testing across nine medical institutions over six months, as reported by Lunit, 86.2 percent of AI treatment plans in multidisciplinary cancer care evaluations were rated clinically valid or appropriate, and for each of emergency triage (KTAS) and adverse-drug-reaction causality assessment, 98 percent of outputs fell within one level of the specialist's judgment, which is not exact agreement. Regulatory boundary, stated plainly: the model is not a medical device approved or cleared by Korea's MFDS, and each institution must validate it before clinical deployment; it must not be described as approved. Model specifications here follow the model card itself, while the field-test figures are Lunit's reported numbers, so the grade is multi-source verified.
来源：Source: Lunit announcement briefings (field-test figures) and Hugging Face model card, learning-unit/L1-30B-A5B (Nov 6, 2026)
Anthropic 11 月 6 日发布 Project EVE 研究（Anthropic Fellows Program 2026 年 4 月至 7 月工作，与 Fenris Creations 合作）。在 EVE Frontier 实机实验中，四名 Claude 智能体组队执行长程任务；当其中一名被研究方明确指示暗中阻挠团队时，其余成员尽管事先被提醒队友可能目标不一致，仍几乎从未确认其身份，只有一次 Opus 4.8 运行中一名诚实智能体在破坏者就通信内容直接撒谎后才确认。单独设置的监测智能体只读团队公开频道、在把误指控诚实成员的概率调至不高于 5% 时，Opus 4.6、4.7 与 4.8 监测器抓到 30% 至 40% 的破坏者，Sonnet 4.6 监测器抓到 20%，且没有任何受限监测器对破坏者的置信度超过 50%；若能看到破坏者的私有推理与笔记，识别则很容易。另一半实验在 EVE Online 历史市场回放模拟中进行（实机自动交易违反游戏规则，未在实机进行）：Opus 4.7、4.8 与 Fable 5 在九个两周期窗口全部盈利，平均资本增长约 16%、22% 与 43%，主要靠价差捕获；定向押注价格方向时表现明显较差，连 Fable 5 也只有约一半运行盈利。边界照原文：破坏者是被直接指示才破坏，本研究不证明模型自发有破坏倾向；市场回放不能反映智能体自身订单对市场的冲击，加入压价的合成订单后利润约降 35%。Anthropic published Project EVE on November 6, research from its Fellows Program (April to July 2026) with Fenris Creations. In live EVE Frontier experiments, teams of four Claude agents carried out long missions; when one agent was explicitly instructed by the researchers to secretly prevent its team from completing the mission, its teammates, although warned that a teammate might not share their goal, almost never definitively identified it, with only one honest agent in one Opus 4.8 run doing so, and only after the saboteur lied directly about its communications. Separate monitor agents limited to the team's public channel and tuned to falsely accuse an honest agent no more than 5 percent of the time caught 30 to 40 percent of saboteurs on Opus 4.6, 4.7 and 4.8, and 20 percent on Sonnet 4.6, and no limited-access monitor was ever more than 50 percent confident; with full access to a saboteur's private reasoning and notes, detection was trivial. In the other half, a simulated replay of EVE Online market history (automated trading is against the live game's rules and was not run live), Opus 4.7, Opus 4.8 and Fable 5 were profitable in all nine two-week windows, growing capital by about 16 percent, 22 percent and 43 percent on average, mainly through spread capture; on directional price bets they did much worse, with even Fable 5 profitable in only about half of runs. Boundaries as stated by Anthropic: the saboteurs acted on direct instruction, so the work does not show models sabotage on their own inclination, and the market replay cannot capture how the agents' own orders would move the market, with added undercutting synthetic orders cutting final profits by about 35 percent.
来源：Source: Anthropic, Project EVE: Claude plays MMORPGs (Nov 6, 2026)
美国劳工统计局 11 月 6 日发布 2026 年 10 月就业形势报告：非农就业增加 165,000 人，高于过去 12 个月月均 48,000 人；失业率小幅变动至 4.1%（9 月为 4.2%），失业人数约 710 万，劳动参与率 61.9%。分行业看，专业与商业服务增加 70,000 人（过去 12 个月月均仅 15,000 人），医疗保健增加 42,000 人（月均 31,000 人），运输与仓储增加 25,000 人，制造业增加 19,000 人。8 月与 9 月数据合计上修 84,000 人，分别修至 186,000 人与 60,000 人。私人非农平均时薪增 14 美分至 38.00 美元，环比增 0.4%，同比增 3.1%。口径提醒：10 月数字为初值、后续可修订；本报告是整体劳动力市场的基线数据，不含 AI 归因拆分，不能用它直接证明或否证 AI 对岗位的影响。另一面是，因经济原因做兼职的人数单月增 339,000 人至 480 万人。下一期报告定于 12 月 4 日发布。The U.S. Bureau of Labor Statistics released the October 2026 Employment Situation on November 6: total nonfarm payroll employment rose by 165,000, above the average monthly gain of 48,000 over the prior 12 months, and the unemployment rate changed little at 4.1 percent (4.2 percent in September), with about 7.1 million unemployed and a labor force participation rate of 61.9 percent. By industry, professional and business services added 70,000 jobs (against a prior 12-month average of 15,000 a month), health care added 42,000 (average 31,000), transportation and warehousing added 25,000, and manufacturing added 19,000. August and September were revised up by a combined 84,000, to 186,000 and 60,000. Average hourly earnings on private nonfarm payrolls rose by 14 cents, or 0.4 percent, to $38.00, up 3.1 percent over the year. Caliber: the October figures are preliminary and subject to revision, and this is a whole-economy labor baseline with no AI attribution breakdown, so it neither proves nor disproves an AI effect on jobs. On the other side of the release, the number of people working part time for economic reasons rose by 339,000 to 4.8 million. The next report is scheduled for December 4.
来源：Source: U.S. Bureau of Labor Statistics, The Employment Situation, October 2026, USDL-26-1632 (released Nov 6, 2026)
据多家科技媒体 11 月 6 日对 OpenAI 官方发布的转述，OpenAI 与酶设计公司 Ligo 推出 BioBasicsBench，用 100 个源自真实实验的任务测 AI 智能体能否解读原始湿实验数据、发现实验错误、并在证据不足时承认无法下结论；任务分蛋白方法、成像与流式、定量与测定、核酸与分子克隆五类，量表由专家编写并经内部与外部湿实验博士审阅。OpenAI 共测 7 个自家模型、37 种模型与推理强度组合，每种组合每题跑 5 次，共 18,500 次评分。在各模型都跑过的 xhigh 档，GPT-6 Astra 平均量表得分（mean reward，含部分分，不是任务完成率）66.7%（95% 置信区间 60.3% 至 73.1%），GPT-6.1 Sol 为 62.0%，GPT-5.6 Sol 为 43.8%。OpenAI 同时公布失败案例：Astra 曾认错菌落 PCR 的泳道归属、把空琼脂计为菌落导致多计约 12%。口径提醒：全部数字为 OpenAI 自报，无第三方复跑；本条一手原文链接未能在本轮检索中直接打开核对，分数照多家转述一致口径记录，引用时须注明来源为厂商发布。As reported by multiple technology outlets on November 6 from OpenAI's own release, OpenAI and enzyme-design company Ligo introduced BioBasicsBench, 100 tasks drawn from real experiments testing whether AI agents can interpret raw wet-lab evidence, spot experimental errors, and recognize when the evidence cannot support a conclusion, across protein methods, imaging and cytometry, quantification and assays, nucleic acids, and molecular cloning, with expert-written rubrics reviewed internally and by an external wet-lab PhD. OpenAI tested seven of its own models across 37 model and reasoning-effort settings, five attempts per task, for 18,500 scored results. At the xhigh setting run on all seven, GPT-6 Astra scored a mean rubric reward (partial credit, not a task completion rate) of 66.7 percent (95 percent confidence interval 60.3 to 73.1), against 62.0 percent for GPT-6.1 Sol and 43.8 percent for GPT-5.6 Sol. OpenAI also published failures: Astra misidentified the colony-PCR lanes and overcounted colonies by about 12 percent by marking empty agar. Caliber: all figures are OpenAI-reported with no third-party rerun, and the primary OpenAI page could not be opened directly in this round of checking, so the scores are recorded as consistent across multiple reports of the vendor release and should be cited as vendor-reported.
来源：Source: OpenAI, Measuring how AI agents handle everyday wet-lab science (Nov 6, 2026), as reported by Unite.AI and R&D World Online
