CREATE WITH AI

AI 创作与商业情报

AI for creative work and business.

关注能用于创作、营销与电商的产品更新，附原始来源。

Product updates for creation, marketing, and commerce, with original sources.

2026.09.10 · LATEST最新一期日报Latest daily issue阅读日报 →Read daily →

最新情报

Latest intelligence

08
AI CreationOpenAI2026-09-09

官方可用性更新

Official availability update

ChatGPT Voice 调整模型选择与每日时长，套餐差异更明确ChatGPT Voice changes model selection and daily limits

OpenAI 9 月 9 日更新 Voice：在需要搜索或更难推理时可使用 GPT-5.6 或 GPT-6 Astra，并沿用文本聊天的模型与推理强度控制。GPT-Live 每日额度改为 Go 3 小时 mini、Plus 3 小时、100 美元 Pro 15 小时、200 美元 Pro 不限；Plus 和 Pro 到限额后不再自动切到 mini。

OpenAI’s September 9 Voice update lets harder search/reasoning use GPT-5.6 or GPT-6 Astra with the text-chat controls. GPT-Live daily allowances are Go: three hours of mini, Plus: three hours, $100 Pro: 15 hours and $200 Pro: unlimited; Plus and Pro no longer automatically fall back to mini after a limit.

编辑观察：把配音、口播校对和客户回访分成独立时段，记录每次任务的时长、人工复听率和是否达到套餐上限；不要把重要交付押在不透明的自动降级上。

Editorial note: Separate narration, read-through and customer-follow-up sessions; record duration, human review rate and plan-limit hits rather than relying on an opaque automatic fallback.

详情、范围与来源Details, scope and sources原始来源 / Source ↗
AI CreationOpenAI2026-09-09

官方协作更新

Official collaboration update

ChatGPT Library 可共享文件夹，资料所有权与撤权需要先想清楚ChatGPT Library sharing adds folders, permissions and an ownership boundary

同一份 9 月 9 日更新加入 Library 文件和文件夹共享：可邀请指定查看者或编辑者，也可在工作区共享，并在对话中使用“Shared with me”资料。

The same September 9 release adds Library file and folder sharing: teams can invite named viewers or editors, share across a workspace, and use material received in Shared with me within a conversation.

编辑观察：为每个品牌项目设一个所有者、资料清单和撤权前导出流程；用一份非敏感素材测试查看、编辑、对话引用和撤权后的实际行为。

Editorial note: Assign an owner, asset inventory and pre-removal export path to each brand project; test view, edit, in-chat reuse and access removal with a non-sensitive asset.

详情、范围与来源Details, scope and sources原始来源 / Source ↗
AI CommerceShopify2026-09-08

官方店面设置更新 · 补读

Official storefront settings update / catch-up

Shopify 把 Meta 列为 AI 渠道，商品默认经 Catalog 共享Shopify lists Meta as an AI channel with Catalog sharing by default

Shopify 9 月 8 日把 Meta 加入后台 Agentic Storefronts 的 AI channels。变更说明称商品会默认通过 Shopify Catalog 与 Meta 共享，商家可管理 Catalog、direct checkout 的访问并查看该渠道表现。

On September 8, Shopify added Meta to the AI channels in Agentic Storefronts. Its note says products are shared with Meta by default through Shopify Catalog, while merchants can manage Catalog and direct-checkout access and view channel performance.

编辑观察：在测试店先核对 Catalog 的产品标题、价格、库存、运费与退货字段；确认哪些商品可被代理发现、哪些结账入口已授权，再查看渠道报表。

Editorial note: In a test store, audit product title, price, inventory, shipping and returns in Catalog; confirm which products are discoverable and which checkout paths are authorized before reading channel reporting.

详情、范围与来源Details, scope and sources原始来源 / Source ↗
AI CommerceSierra2026-09-08

开源评测 · 非产品能力承诺

Open benchmark / not a product-capability claim

Hyper-τ-bench 把“让代理造客服代理”变成可验收问题Hyper-τ-bench evaluates agents that build customer-service agents

Sierra 9 月 8 日开源 hyper-τ-bench：开发代理要从模拟企业的记录中恢复需求、构建客服代理，并在未见过的模拟流量上验收，而不只看能否生成代码。

Sierra open-sourced hyper-τ-bench on September 8. A developer agent must recover requirements from simulated business records, build a support agent, and be assessed on unseen simulated traffic—not merely whether it generates code.

编辑观察：把自家客服代理的验收拆成需求回收、工具调用、越权处理、人工交接和盲测五项；上线前用未参与编写答案的人维护一组隐藏案例。

Editorial note: Split a support-agent acceptance plan into requirement recovery, tool use, authorization handling, human handoff and blind tests; maintain hidden cases with people who did not author the answers.

详情、范围与来源Details, scope and sources原始来源 / Source ↗
AI CommerceXinhua / Meituan2026-09-09

服贸会现场报道 · 平台自报

CIFTIS report / platform-reported

美团在服贸会展示 CatPaw 与商家智能体，作为经营工作台信号Meituan showcases CatPaw and merchant agents at CIFTIS

新华网 9 月 9 日报道，美团在服贸会展示 CatPaw AI 经营工作台及“智能掌柜”等应用，所述场景包括评价分析、经营诊断、门店巡检与按店铺需求配置代理。

A September 9 Xinhua report says Meituan presented its CatPaw operating workbench and merchant-agent applications at CIFTIS, including review analysis, business diagnosis, store inspection and store-specific configuration.

编辑观察：把门店巡检先限定为可核验字段：缺货、评价主题、活动到期和食品安全记录；让人复核建议，再记录采纳率与错误类型。

Editorial note: Start store inspection with verifiable fields—stockouts, review themes, expiring promotions and food-safety records; have people review recommendations and track adoption and error types.

详情、范围与来源Details, scope and sources原始来源 / Source ↗
AI MarketingXinhua / Taotian2026-09-09

服贸会现场报道 · 平台自报

CIFTIS report / platform-reported

淘天在服贸会展示 AI 购物、创意与售后组合，但数字需归因Taotian showcases AI shopping, creative and support tools at CIFTIS

新华社客户端 9 月 9 日报道，淘天在服贸会集中展示面向导购、商家经营、广告创意、采购和售后的一组 AI 应用，包括 AI 购物助手、AI 店小蜜与阿里妈妈创意/定向产品。

A September 9 Xinhua client report says Taotian presented a set of AI applications for shopping guidance, merchant operations, advertising creative, procurement and support, including its shopping assistant, Dianxiaomi and Alimama products.

编辑观察：营销团队把“素材生成、定向、导购、售后”分开建指标；一次实验只改变一个环节，并保留人工兜底和原始订单口径。

Editorial note: Measure creative generation, targeting, shopping guidance and support separately; change one stage per experiment, retain human fallback and preserve the original order-definition metric.

详情、范围与来源Details, scope and sources原始来源 / Source ↗
AI CommerceAI World2026-09-10

AI World 工作台 · 今日原创

AI World workbench / original today

工作台：为代理式商品发现做一次字段与授权检查Workbench: check product fields and permissions for agentic discovery

基于 Shopify 将 Meta 列为 AI channel 的设置变化，本工作台把代理式发现当成数据与权限检查，而不是新增销售渠道的结论：每个商品需要一个可验证的标题、价格、库存、运费、退货和结账授权状态。

Based on Shopify’s Meta AI-channel setting, this workbench treats agentic discovery as a data-and-permission audit, not as a conclusion that a new sales channel will perform. Each product needs verifiable title, price, inventory, shipping, returns and checkout-authorization status.

编辑观察：导出 20 个 SKU，逐条比较站内页和 Catalog 字段；把缺失、相互矛盾和需人工确认的字段标色，并在测试环境复核 direct checkout 的开关。

Editorial note: Export 20 SKUs and compare storefront and Catalog fields; flag missing, conflicting and human-confirmation fields, then verify direct-checkout settings in a test environment.

详情、范围与来源Details, scope and sources原始来源 / Source ↗
AI CommerceAI World2026-09-10

AI World 工作台 · 今日原创

AI World workbench / original today

工作台：客服代理上线前，用隐藏案例验收而非只看演示Workbench: accept a support agent with hidden cases, not a demo alone

借鉴 hyper-τ-bench 的“未见流量验收”思路，本工作台给电商客服代理准备五类隐藏案例：退款例外、库存冲突、身份/权限、模糊描述和必须转人工的高风险请求。

Drawing on hyper-τ-bench’s held-out-traffic idea, this workbench prepares five hidden-case classes for a commerce support agent: refund exceptions, inventory conflicts, identity/authorization, ambiguous requests and high-risk cases that must reach a human.

编辑观察：每类至少写 10 条隐藏测试，定义正确答复、必须升级条件和禁止动作；由未参与提示词编写的人打分，公布失败样本后再改流程。

Editorial note: Write at least 10 hidden tests per class with expected answer, mandatory escalation and prohibited actions; have non-authors score them, then revise only after reviewing failed samples.

详情、范围与来源Details, scope and sources原始来源 / Source ↗
2026-W36 · 2026.08.31—09.06
WEEKLY

从模型发布到实际交付

From model release to usable work

用一件商品，验证资料、卖点、分镜与页面能否连成一次合格交付。

Use one product to test research, copy, storyboarding, and page delivery.

阅读专题 →Read feature →