1. GPT-5.6 enters Kiro / GPT-5.6 进入 Kiro
OpenAI says GPT-5.6 is now available in Kiro across software planning, building, review, and testing. The release emphasizes developer value: GPT-5.6 Terra reportedly completed Terminal-Bench 2.1 tasks in Kiro at roughly 82% lower cost.
OpenAI 将 GPT-5.6 接入 Kiro，覆盖规划、构建、审查和测试。关键并非又多了一个模型入口，而是模型竞争开始用“完成真实开发任务的单位成本”衡量。
链接：https://openai.com/index/gpt-5-6-in-kiro
2. Stampli cuts release work by 68% / Stampli 将发布工时缩短 68%
Stampli reports that Codex and ChatGPT Work compressed release-production work from weeks to days, reducing the required labor by 68%. This is a rare quantified example of AI workflow ROI rather than a generic productivity claim.
Stampli 用 Codex 与 ChatGPT Work 把数周的发布制作压缩到数天，工时降低 68%。这类可核对的流程指标，比“效率提升”式口号更能推动企业采购。
链接：https://openai.com/index/stampli
3. Sheets becomes an application canvas / Google Sheets 正变成应用画布
Google introduced a Sheets canvas that can turn prompts into interactive dashboards, trackers, and layouts. Spreadsheet incumbents are expanding from record-keeping into lightweight application generation.
Google 让用户通过自然语言在 Sheets 中生成仪表盘、追踪器和布局。表格正在从 system of record 进化为轻应用执行环境，对简单垂直 SaaS 构成直接压力。
链接：https://blog.google/products-and-platforms/products/workspace/sheets-canvas-for-google-sheets-spreadsheets/
4. Canvases make agents visible and steerable / Canvas 让 agent 可见、可控
GitHub argues that chat streams are a poor interface for long-running agent work. A canvas can expose state, accept human intervention, and help control cost—turning supervision into a product surface.
GitHub 指出，聊天滚动流无法承载长期 agent 任务；canvas 能呈现状态、允许人工干预并控制成本。agent 产品的关键界面正在从“对话框”转向“可监督的工作空间”。
链接：https://github.blog/ai-and-ml/github-copilot/how-canvases-make-agentic-workflows-visible-steerable-and-cost-efficient/
5. SaaS must become an agent harness / SaaS 必须进化为 agent harness
Garry Tan argues that systems of record must evolve into AI harnesses or risk replacement by agents. Related builder signals point to gateways, staged evals, permissions, and observability as the new operational layer.
传统 SaaS 不能只靠数据保管形成壁垒，还要把上下文、权限与工作流开放为 agent 可执行、可评估、可审计的 harness。AI Gateway 和分阶段 eval 将成为这一转型的基础设施。
链接：https://x.com/garrytan/status/2091742825042030681
我的判断
今天五条信号汇合为一条主线：AI 正从“能力展示”进入“可信执行”。模型价格下降会扩大需求，但企业真正愿意迁移关键流程，取决于任务完成成本、可观测性、分阶段 eval、权限控制和人工接管。
对 opcpay.org 读者的意义
支付与金融 SaaS 天然具有高权限、高审计和高错误成本，更需要 agent control plane。创业机会不仅在新模型应用，也在工作流 ROI 计量、权限与审计、模型路由、异常回滚，以及把既有 system of record 改造成 agent harness。