Shutdown Winter: OpenAI, Google and ByteDance Pull the Plug on Their AI Bot Builders

What this is about

In late 2023, the AI industry made a promise: anyone can build their own bot, no coding required. Builder tools like OpenAI's Custom GPTs, ByteDance's Coze and Google's Opal were how you did it. You typed instructions into a text box, uploaded a few PDFs and wired boxes together on a visual canvas into a workflow. With this so-called no-code approach, you click a program together instead of writing it.

In fall 2026, these same vendors are clearing out their builders within a few weeks of each other. Our thesis: these aren't isolated product decisions. A whole tooling layer is being pushed aside, by AI agents that write their own tools. If you run bots or workflows on these platforms, you need to act now, because the first deadlines expire in a few days.

The shutdown calendar

The OpenAI dates for Agent Builder, Evals and Prompts are listed on the official deprecations page and were announced on June 3, 2026. Agent Builder had only been introduced in fall 2025, so it was deprecated after roughly eight months. The GPT dates come from OpenAI's migration FAQ, which says timelines and features may vary by workspace. For Vertex AI, Google's release notes from May 26 say the Extensions are deprecated and will be "shut down after November 26, 2026."

Coze: market leader, and out anyway

The shift is clearest at Coze, known as Kouzi in Chinese. According to 36Kr, the platform had a 19.3 percent share in 2025, making it number one among China's cloud platforms for agent development. According to the official discontinuation notice from October 29, ByteDance will stop accepting new customers for the Enterprise Standard and Flagship plans from November 30. It's over for good at the end of 2027. As recently as July, Coze had raised the Standard plan from 498 to 980 yuan a month. How conversations, memory, files and workflows can be exported, Coze says it will clarify only "in a later notice."

You're not completely stuck, though. A guide on smzdm checked what already works today:

• Low-code workflows can be exported as a ZIP with a YAML description. YAML is a human-readable text format. But the description follows a Coze-specific schema and can only be imported back into Coze.
• Knowledge bases can't be exported. You need the original PDFs and documents.
• Conversation logs can be downloaded as a CSV file, up to 5,000 messages per export. But they're only kept for 90 days (Standard) or 180 days (Flagship). If you don't export now, you lose them.
• Long-term memory and chats in Coze 3.0 have no export at all so far.

According to Sina Tech, Coze isn't alone. Tencent's Yuanqi goes offline on November 9, and Baidu's Miaoda was folded into Baidu Dazi. An explicitly incomplete list compiled by Tech星球 counts 38 discontinued or merged AI products from Chinese companies, 25 of them in the second half of 2026 alone.

Why the builders are disappearing

36Kr sums up the argument neatly: drag-and-drop flowcharts were a crutch for weak models. Once a model can break a task down into steps on its own, the canvas becomes a pointless middle step. The goal of large models, it says, is "to flatten the canvas itself." A former Coze employee quoted by woshipm says essentially the same thing.

Fall 2026 shows what's replacing the builders:

• OpenAI is replacing GPTs with plugins that run in ChatGPT and in the coding agent Codex. With ChatGPT Work and Sites, the assistant builds web apps from company data on its own.
• Meta has let its assistant Muse build custom mini-apps on request via the Spaces feature since November 6.
• DeepSeek unveiled Creator Mode for Harness on November 2: describe what you want, and the agent programs a plugin for itself, installs it and tests it.
• Google is sending Firebase Studio and Jules users to Antigravity, its development environment where AI agents do most of the work.
• ByteDance folded the Coze team into the Doubao division in August and launched Doubao Work the next day. Microsoft is consolidating natural-language app building in Copilot Cowork.

Not all of this is new. In 2023, OpenAI replaced the original ChatGPT plugins with GPTs. Gizmodo reported at the time on annoyed developers posting "RIP Plugins." Three years later, it's flipping back: GPTs are giving way to plugins again, except that today's plugins are built differently.

What the move costs you

The firsthand accounts are sobering. In the OpenAI developer forum, one user calls the shutdown "not a solid business strategy." Another had built client projects on Agent Builder and ChatKit, OpenAI's building block for chat interfaces. Until now, OpenAI provided the backend, the server side. Now he has to run it himself, which by his own account makes his projects "economically unviable." A third reports cold starts of around 30 seconds, the wait on the first request, on a self-hosted FastAPI server. An always-on server costs him 15 to 30 dollars a month. OpenAI hasn't replied in the thread.

Blogger Hidekazu Konishi read the documentation closely. His verdict: Agent Builder's code export is a starting point, not a lossless translation. You have to double-check triggers, tools and permissions yourself. The no-code alternative, Workspace Agents, requires a Business, Enterprise or Edu workspace. Whether the export still works after November 30 isn't documented. He found no bulk export for individual evals and prompts.

With GPTs, too, OpenAI's own warning applies: a migrated plugin may respond differently. In Gemini, according to The Decoder (mixed-news), you can activate at most 100 Skills at once. Sharing and Drive files are supposed to follow only "in the coming weeks."

What has survived every shutdown

Put the cases side by side and a clear pattern emerges:

Skills and MCP are so resilient because of where they come from. Both are open standards. Anthropic published the Skill format as an open specification in December 2025. MCP, the Model Context Protocol, is a kind of universal plug that lets AI models access tools and data sources. Since December 9, 2025, it has belonged to the Linux Foundation's Agentic AI Foundation, whose members include OpenAI, Google and Microsoft. According to OpenAI's plugin documentation, a plugin is also just a folder with a plugin.json manifest, an mcp.json and a skills/ subfolder. OpenAI-specific settings live separately in their own extension field.

Is there a countermovement?

Open-source builders like Dify, n8n or the self-hosted Coze Studio (Apache 2.0 license, around 21,800 GitHub stars) pitch vendor independence. But we found no solid numbers on users switching over. The Chinese comparison articles often come from vendors like Tencent Cloud recommending their own product. And according to 36Kr, the whole market is small: the Chinese segment Coze led was worth about 250 million yuan.

The bottom line

Our take: the builder era isn't ending because the idea was bad. The models have taken over the work you used to need the boxes for. If you want a bot today, you describe it, and the agent writes the Skill, plugin or mini-app itself. That's more convenient, but it doesn't solve the underlying problem. Meta's Spaces can't be exported either, and DeepSeek itself says Harness hasn't gone through a security review yet. So new walled gardens are emerging that will be just as hard to get out of later.

The lesson from this fall is simple: keep your work in formats you own, meaning text, code, Skills and MCP. If you're affected right now, do it in this order:

1. First, back up data with an expiry date, such as Coze logs or Evals results.
2. Next, save original documents and prompts as plain text.
3. Only then migrate, and use fixed test inputs to check that the new bot still does what the old one could.

Still open: when Coze will publish its export process, and whether OpenAI's code export will still work after November 30.