
OpenAI 在 DevDay 2026 发布了 20 多项新东西,核心是把常驻智能体 dots 放进普通人的订阅;本期逐项核对它公布的模型与价格、台上那句「第一个 AI 研究实习生」里经得起对的部分,以及同一天写进三份文档、却没有出现在台上 68 分钟里的安全内容。
"The event will offer our best indication yet of how seriously OpenAI is taking the safety concerns about its technology, and whether the company plans to actually "pace" its advancements, like Altman said it was considering to do."
| 方向 | 台上给出的东西 |
|---|---|
| 给个人用 | dots(常驻智能体)、ChatGPT Space(团队协作空间)、Pages、团队任务、共享档案 |
| 模型与速度 | GPT-6.1 Sol(新模型)、Astra 与 Sol 的 Ultrafast 加速档 |
| 给开发者 | Codex 上云、Codex Security Cloud、Agents API 支持电脑操作、与 AWS 合作的 Bedrock Managed Agents、插件扩展、MCP Events |
| 分发与付费 | Sign in with ChatGPT(16 家伙伴)、OpenAI Marketplace(首批 32 家)、Pro 500 档 |
| 数据与隐私 | Zero Data Retention with Private Safety Processing、Private Inference 预览 |
"Dots are remarkably capable, always-on agents built to handle everything."8

"We tracked how often a model can carry out a task in research. It failed the vast majority of times on a task that took a day or longer, but as of July our models have taken on over a third of day long research tasks with no intervention required."2
a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days——能在人类指导下完成定义明确的研究任务,其中包括熟练研究者要做好几天的任务。11"In the last 6 months, over half of successful 4-8 hour tasks involved 1 or more interventions."11

Measurements above 16 hrs are unreliable with our current task suite,并且专门回答过一个误解:8 小时的时间视野,不等于 AI 能做完一个高语境人类专家一天的工作。12 截至该页更新,GPT-6 系列的时间视野都还没有公布。Claude is not operating fully autonomously for any measured subset of AI R&D work。13The agents completed all of the engineering without human help, yet could not make substantial progress towards answering the research questions.——工程部分它们全都自己做完了,但没有在回答研究问题上取得实质进展。14Dots can still make mistakes, and we'll keep improving these protections as we learn from real use.——dots 仍然会犯错。
We observed no attempts to bypass an automated safety reviewer, matching GPT-6 Astra and GPT-6 Sol.9our most capable and aligned model,并写明We've trained Astra to refuse harmful requests, including requests involving biological or cybersecurity misuse.——最能干、也最对齐的那一档,训练它拒绝对生物和网络安全的滥用请求。16Desks in our offices, access badges, and company laptops. 评估者有权公开风险与事故的结论,without editorial control by Anthropic——不受 Anthropic 的编辑控制。六天后,Anthropic 宣布已与埃森哲签约推进这件事。1819Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release. 先自愿送审,最长提前 30 天;协议被证明有效之后再转成上市前强制通过。20"If you build a product or a service and you're not confident in its functionality, capability, or safety, then don't release it. … The market forces are already there. We don't need any new laws. We don't need any new regulations."21
OpenAI is committed to supporting independent assessments with deep levels of access across training, evaluation, and deployment.——承诺支持带深度访问权的独立评估,必要时access on company-managed devices or premises may be appropriate。23| 谁 | 主张的机制 | 谁来判断、能不能公开 | 时点 |
|---|---|---|---|
| Anthropic(Amodei,9 月 12 日) | 第三方评估者嵌入公司,员工级访问权限 | 放行决定仍由公司作出;评估者独立核查并公开发表,不受公司编辑控制 | 已单方面承诺,9 月 18 日与埃森哲签约推进 |
| Google DeepMind(Hassabis,7 月 14 日) | 类 FINRA 的行业标准机构 | 标准机构做测试;协议成熟后成为上市前提 | 先自愿送审,最长提前 30 天 |
| 英伟达(黄仁勋,9 月 15 日与 28 日) | 不新增机制,当作工程问题交给市场 | 公司自己 | 现在 |
| OpenAI(9 月 22 日、28 日两份文件) | 深度访问的独立评估+训练前安全案例 | 文件写审计者应有足够权限;同一份文件写安全案例由高级管理层评审、每人可否决 | 两份都是主张,未给出启用时间 |
more than 99 percent reliability、cut the time to first token by 45 percent、the most reliable API available on the market today——在 DevDay 回顾页、GPT-6.1 Sol 发布页和开发者 changelog 里都没有对应表述。2our collective 1.2B weekly users——周活跃 12 亿,不是注册数,也不是月活。7 这个口径下没有第三方独立读数;第三方给出的数字是另一回事:Sensor Tower 测到 ChatGPT 移动应用在 2026 年 5 月达到 10 亿月活27,a16z 的榜单在 2026 年 3 月记的是 9 亿周活28——量的是别的范围、别的时点,和 12 亿放不到同一把尺子上。openai.com
youtube.com
openai.com
openai.com
openai.com
openai.com
alignment.openai.com
darioamodei.com
anthropic.com
demishassabis.substack.com
status.openai.com
llmlatency.dev
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.