从模型发布到实际交付
From model release to usable work
以 GPT-6 的能力与开放进展为切口，用一件商品，把资料、卖点、分镜和页面连成一次可检查的工作。
Use GPT-6 capability and rollout updates to examine one practical task: take a product from source material through copy and storyboarding to a reviewable page.
发布于 2026.09.07 · 已补充新手用法与 3D 案例拆解 · 非实测报告
Published 2026.09.07 · Updated with beginner steps and 3D case breakdowns · Not a test report
Start here · Case breakdowns · Try a project
第一次用 GPT-6，从这里开始
Your first task with GPT-6
先选工作入口：问问题、讨论思路，用 Chat；需要一份可检查的文件或成品，用 ChatGPT Work；要制作网页、游戏或修改项目，用 Codex 也可以。入口和工具是否可用，取决于你的账号与工作区设置。GPT-6 的开放范围见下文，不能仅凭选择了 Work 就判断正在使用 Astra。
Choose a working mode: Chat for questions and discussion, ChatGPT Work for a reviewable deliverable, or Codex for a website, game, or project changes. Available modes and tools depend on your account and workspace. Check Astra access below; choosing Work alone does not establish which model you are using.
接着给它一份材料和一个小目标。例如上传一篇文章，要求改成初学者教程。写清楚给谁看、要解决什么、最后交什么；第一版不合适，就在同一任务中继续指出具体问题，不需要重新开始。
Then provide material and a small goal. For example, attach an article and ask for a beginner tutorial. State the audience, purpose, and deliverable. Refine the first result in the same task with specific feedback.
可以这样说：“把这篇文章改成初学者能照着做的教程。保留有来源的事实，合并重复内容，每一步给一个例子。交付完整成稿；影响结果的关键信息缺失时先问我。”
Try this: “Turn this article into a tutorial a beginner can follow. Preserve sourced facts, merge repetition, and give an example for each step. Deliver the complete draft. Ask about missing information that would materially change the result.”
看到结果后继续说：“第二步仍然抽象，写到我能照着操作。”最后打开成稿或预览，核对它是否真的满足要求。让模型自查有帮助，但不能代替你查看最终作品。
Follow up with: “Step two is still abstract; make it actionable.” Open the draft or preview and check it against the brief. Model self-review helps, but does not replace looking at the final work.
本周信号：先确认自己能用什么
This week: establish what you can access
9 月 4 日的模型介绍与 9 月 6 日的开放进展，是同一条变化的两个阶段。本期把它们合并来看：新能力值得关注，但能否进入工作，还取决于产品入口与账号权限。
The September 4 model brief and September 6 rollout update describe two stages of one development. Read together, they raise a practical question: which capabilities can your product and account actually use?
官方事实：GPT-6 Pro 由 GPT-6 Astra 驱动，正向 ChatGPT 的 Pro、Business 与 Enterprise 分批开放；Plus 的 Astra 入口在 ChatGPT Work 和 Codex，也在分批开放。Enterprise 还取决于工作区权限。
Official fact: GPT-6 Pro is powered by GPT-6 Astra and is rolling out to ChatGPT Pro, Business, and Enterprise. Plus receives Astra in ChatGPT Work and Codex during rollout. Enterprise access also depends on workspace permissions.
因此，开始前先记录实际可见的模型、产品入口和工具权限。不要把“官方已经宣布”直接写成“团队今天都能使用”。
Before starting, record the model, product, and tools actually available to your account. An announcement does not establish availability for every team member.
能力与交付之间，还隔着工具和验收
Tools and acceptance sit between capability and delivery
官方事实：Astra 的模型说明列出研究、代码、电脑操作和文档制作等用途；它接收文字与图片，输出文字。图片生成是可调用的工具，不能据此把它描述成原生视频或音频生成模型。
Official fact: Astra documentation lists research, coding, computer use, and document creation. It accepts text and images and produces text. Image generation is a supported tool; this does not make Astra a native video or audio generator.
编辑判断：对创作者，更有意义的验证是能否把资料整理、内容构思与交付检查接起来。单张图好看，不能证明商品资料准确；页面生成了，也不能证明手机上可读。每一步都需要明确的合格标准。
Editorial view: creators should test whether research, content planning, and delivery checks work together. An attractive image does not establish product accuracy. A generated page does not establish mobile readability. Each step needs explicit acceptance criteria.
那些 3D 和游戏案例，是怎么做出来的？
How do the 3D scenes and games get made?
“三个三”的十个案例串帖提供了另一种观察角度：Blender 场景与建筑、浏览器游戏、Unreal 世界、KiCad 电路板布局，以及 3D 效果图工具。它们展示的是模型与软件组合的工作，不能统一理解为在聊天框输入一句话就直接得到全部成品。
The ten-example thread by 3three_AI points to Blender scenes and architecture, browser games, an Unreal world, KiCad board layout, and a 3D mockup tool. These involve models working with software; they are not interchangeable examples of a chat message producing a finished asset.
串帖引用的 Higgsfield 展示明确描述了“场景描述 → Astra 生成代码 → Blender 建模 → Cycles 渲染”的过程。另一个曼哈顿展示注明用了约一周。本文读到了这些作者说明，但没有复现十个演示，也不把比较视频当成统一条件下的模型评测。
The quoted Higgsfield demonstration describes scene direction, Astra-generated code, Blender modeling, and Cycles rendering. The Manhattan example describes work over roughly a week. We read those author accounts; we have not reproduced all ten demos or treated their comparisons as controlled benchmarks.
社区案例索引 / Community case index ↗
官方建筑案例提供了更完整的过程：先确认布局，再建模，查看渲染图并修正问题，然后制作镜头运动。官方游戏案例也记录了反复试玩、定位问题与重新检查。对新手可借鉴的是这套迭代方法，而非“一次成功”的宣传语。
OpenAI’s architecture walkthrough documents layout review, modeling, render inspection, corrections, and camera work. Its game walkthrough also documents repeated play, diagnosis, and testing. The useful beginner lesson is the iteration process.
想试这种能力，先做一个小型展示工具
Start with a small interactive showcase
建议起点：可旋转的商品展示页。给一张商品参考图，要求先用简单几何体做交互原型，支持拖动旋转、切换三个颜色和重置视角。第一轮只验收这些动作，再逐步改外观。这样能把“能生成画面”推进到“真的能操作”。
Suggested starting point: a rotatable product page. Provide a reference image and ask for a simple geometric prototype with drag rotation, three color choices, and a camera reset. Check those interactions first, then refine the appearance.
可直接使用的任务：“请做一个能在浏览器打开的商品 3D 展示页。参考我提供的图片，先用简单几何体搭建；支持拖动旋转、三种颜色和重置视角，兼容手机。请运行并检查这些交互，交付预览和源文件。未能从图片确认的尺寸标为估计值。”
Task to try: “Build a browser-based 3D product showcase from my reference image, starting with simple geometry. Add drag rotation, three colors, and a camera reset, including mobile support. Run it, check those interactions, and deliver a preview and source files. Label dimensions inferred from the image as estimates.”
如果想要可编辑的 3D 场景，再选择 Blender 路线：先确认电脑已安装 Blender，且当前工作环境能使用它；从一张桌子和一个商品开始，要求交付 .blend 文件与预览图。仅有参考照片时，得到的是外观近似，不是精确的尺寸重建。这里的练习是编辑建议，未附实测耗时或成功率。
For an editable 3D scene, take the Blender route: confirm Blender is installed and accessible to the working environment, then begin with one table and one product. Request a .blend file and preview renders. Photos alone support an approximation, not exact dimensional reconstruction. These are suggested exercises, without claimed timings or success rates.
用同一件商品，跑完一段真实工作
Take one product through a complete task
建议尝试，非本刊实测：选一件资料齐全的商品，提供规格表、三张授权图片、目标受众与不能使用的宣传表述。所有环节沿用同一份资料，不让模型在中途补造参数。
Suggested exercise, not a reported test: choose a product with a verified specification sheet, three authorized images, a target audience, and prohibited claims. Use the same source pack throughout; do not allow invented specifications.
第一步，整理三条卖点，每条标明对应资料；缺少证据的卖点先留空。第二步，写六镜头分镜，逐镜说明画面、文案与所需素材；没有生成的视频就保留为分镜，不称为成片。
First, draft three benefits and connect each to its evidence; leave unsupported benefits blank. Second, prepare a six-shot storyboard with visuals, copy, and required assets. A storyboard remains a storyboard until a video is produced.
第三步，把确认过的内容放进商品说明页面。在手机与电脑上打开，检查文字、图片和按钮。若更换工具做对比，继续使用相同资料与验收标准，记录返工而不只看第一版。
Third, place the approved content in a product page. Open it on mobile and desktop and check text, imagery, and buttons. When comparing tools, keep inputs and acceptance criteria constant and record revisions, not just the first output.
验收卡：什么结果才值得继续用
Acceptance card: decide whether to keep using it
资料准确：尺寸、材料与功能逐项对照原始资料；每条卖点有依据。视觉一致：商品外观和关键细节没有在不同镜头里改变。阅读可用：390px 手机宽度与桌面都没有重叠或截断，链接与按钮去向正确。
Accuracy: verify dimensions, materials, and functions against the source pack and support every benefit. Consistency: preserve product appearance and key details across shots. Usability: ensure readable layouts at 390px and desktop widths, with working links and buttons.
记录总耗时、人工修改次数、仍未解决的问题。若只是输出更快、却把更多时间花在改错上，就还没有证明它改善了这段工作。没有真实运行记录时，不填写效率提升、成功率或转化效果。
Record total time, manual revisions, and unresolved issues. Faster output does not establish a better workflow if corrections consume the time saved. Without actual runs, do not report efficiency gains, success rates, or conversion results.
本期建议：先完成这一个小任务，再决定扩大使用。账号尚未开放时，可以先整理输入资料和验收表；不要为尚未验证的能力承诺交付日期。
This issue’s recommendation: finish one bounded task before expanding use. If access has not arrived, prepare the source pack and acceptance sheet first; do not commit delivery dates based on unverified capability.