SANA-WM open-sources a 2.6B world model for one-minute 720p video / SANA-WM 开源 26 亿参数世界模型
English: NVIDIA Labs released SANA-WM, a 2.6B open-source world model capable of generating one-minute 720p videos. The important signal is not only better video generation, but the return of world models as a serious architecture for representing and predicting dynamic environments.
中文：NVIDIA 实验室发布 SANA-WM，一个 26 亿参数的开源 world model，可以生成 1 分钟 720p 视频。真正重要的不是“视频生成又进步了”，而是 world model 重新回到主线，成为理解和预测动态世界的一条关键路线。
链接：https://nvlabs.github.io/Sana/WM/
Databricks brings GPT-5.5 to enterprise agent workflows / Databricks 将 GPT-5.5 接入企业 agent 工作流
English: Databricks is using GPT-5.5 for enterprise agent workflows after the model set a new state of the art on the OfficeQA Pro benchmark. This is a sign that large models are moving from demos and chat interfaces into real enterprise workflow execution.
中文：Databricks 把 GPT-5.5 用进企业 agent workflows，并提到该模型在 OfficeQA Pro 基准上达到新的 SOTA。这说明大模型正在从演示和聊天界面，进入真实企业流程的执行层。
链接：https://openai.com/index/databricks
OpenAI expands Codex use cases for operations and data science / OpenAI 将 Codex 推向业务运营与数据科学团队
English: OpenAI published examples of how business operations and data science teams can use Codex to produce initiative briefs, leadership packets, KPI memos, root-cause analyses, and dashboard specs from real work inputs. Codex is being positioned as a cross-functional work system, not only a coding assistant.
中文：OpenAI 发布了业务运营团队和数据科学团队使用 Codex 的场景，包括 initiative brief、领导决策包、KPI memo、root-cause analysis 和 dashboard spec。Codex 正被定位成跨职能工作系统，而不只是 coding assistant。
链接：https://openai.com/academy/codex-for-work/how-business-operations-teams-use-codex
Builders are converging on agent infrastructure and maintenance / Builder 共识转向 agent 基础设施与长期维护
English: Swyx highlighted national-scale MCP gateway thinking in Singapore; Guillermo Rauch framed the next developer advantage as agent management plus deep fundamentals; Aaron Levie argued that AI deployment requires forward deployed engineering because models and workflows keep changing; Dan Shipper suggested a strong company-level super-agent may be more practical than one personal agent per employee.
中文：Swyx 提到新加坡正在从国家尺度思考 MCP gateway；Guillermo Rauch 认为下一阶段开发者优势来自 agent management 加工程基本功；Aaron Levie 判断 AI 交付需要 FDE 式持续部署；Dan Shipper 则认为公司级 super-agent 可能比每人一个 personal agent 更现实。
链接：https://x.com/swyx/status/2055470634331750588
ChatGPT Finances exposes the trust boundary of AI in financial products / ChatGPT Finances 暴露金融 AI 的信任边界
English: Peter Yang saw ChatGPT Finances as a promising product direction, but immediately turned off model-improvement sharing because he did not want financial data used for training or targeted ads. He also noted that AI still struggles with transaction classification. In finance, privacy defaults and mundane accuracy matter more than impressive demos.
中文：Peter Yang 认为 ChatGPT Finances 是有价值的产品方向，但马上关闭了“用于改进模型”的数据分享，因为他不希望财务数据进入训练或广告定向。他还提到 AI 仍会把交易分类搞错。金融场景里，隐私默认设置和基础准确性比炫技 demo 更重要。
链接：https://x.com/petergyang/status/2055450577094738018
我的判断
今天的核心变化可以概括为一句话：AI SaaS 正在从“模型能力竞争”进入“可信执行系统竞争”。
世界模型解决的是 AI 对动态环境的理解和预测能力；GPT-5.5、Codex、Databricks 解决的是企业流程执行能力；MCP gateway、Vercel agent deployment、OpenClaw 多 Codex 实验解决的是工具、权限、部署和维护问题。三条线最终会汇合到同一个问题：agent 如何在真实组织里长期、可靠、可控地工作。
对 opcpay.org 读者的意义
对 SaaS 创业者来说，下一阶段产品设计不能只问“能不能接入更强模型”，而要问三个更具体的问题：
- 这个 agent 能接入哪些真实业务系统？
- 它如何处理权限、隐私、审计和错误责任？
- 它是否需要 forward deployed engineering 式的持续维护？
如果答案不清楚，产品就很容易停留在 demo。如果答案清楚，就有机会成为客户组织里的新工作层。