Article

Move OpenAI Agent Builder workflows before November 30

For teams using OpenAI Agent Builder, compare SDK and ChatGPT migration options, access requirements and usage costs before the shutdown.

Editorial illustration for Move OpenAI Agent Builder workflows before November 30: an arrow connects an old service to its replacement. Not documentary evidence.

Teams using OpenAI Agent Builder need to move their workflows before its scheduled November 30, 2026 shutdown. The two documented routes are to run exported code in an application with the Agents SDK, or rebuild it as a ChatGPT Workspace Agent. Neither route guarantees unchanged behavior.

OpenAI gave notice on June 3; this is an approaching deadline, not a newly announced closure. The choice is where the workflow will run: your application or a ChatGPT workspace.

Export the workflow while Agent Builder is available

For each workflow, open Code in the top navigation, choose Agents SDK, select TypeScript or Python, and copy the complete export. OpenAI’s migration guide requires access to the original workflow. Keep a copy alongside a record of its owner, connected tools and expected results.

Choose by where the result must go

A useful way to choose between the application-owned SDK and shared workspace agents is to start with who needs the result and how they receive it:

Your requirement

Starting point

Responsibility that remains

Your website or application needs a direct answer and explicit process rules.

Agents SDK in your application.

Operate the backend and validate its tools, storage, permissions and approvals.

Colleagues can do the work inside an eligible ChatGPT workspace.

Recreate a Workspace Agent.

Configure connections and schedules, then validate the rebuilt process.

Workspace Agents can be started by other software. However, their trigger API queues a run rather than returning the agent’s answer. Beta status polling can report progress or completion, but the documentation says the response itself cannot currently be retrieved through that API.

That distinction matters for migration: starting an internal task is different from supplying a live answer to a customer in your app. Do not treat an accepted trigger as a replacement for your existing response channel.

Keep ChatKit, replace its hosted workflow

ChatKit is not retiring. What is ending is the Agent Builder workflow behind an Agent Builder-hosted integration. OpenAI directs migrations to a custom server-side implementation, so retaining the chat interface does not remove the backend work.

The custom ChatKit guide calls for persistent conversation storage, file handling if uploads are supported, and permission enforcement using the signed-in user’s identity. Decide separately how existing records will be retained or moved; a familiar-looking chat window does not establish that its history survived.

With the Agents SDK, your team owns deployment and application integrations while the SDK handles the agent loop. This is distinct from OpenAI’s separately managed Agents API. Running the SDK in your application does not mean model processing happens locally.

Check workspace access and fixed process rules

The workspace migration route requires ChatGPT Business, Enterprise or Edu, agent access, and permission to create agents. A plan name alone does not establish those permissions.

Start a new workspace-agent draft and provide the exported code as rebuilding input. Review the proposed instructions and capabilities before creating the agent. OpenAI warns that workflows relying on fixed, predictable steps may not transfer faithfully.

For example, if a process must always check an approval before changing a record, treat that sequence as a requirement to verify—not merely something to mention in the prompt. The broader choice between a chatbot, workflow and agent is useful when deciding which steps should remain fixed.

Review whose account each connected app uses. OpenAI’s workspace guidance distinguishes each runner’s own account from an agent-owned shared connection. App write actions ask for approval by default, although settings vary. Sharing an agent therefore needs two separate decisions: who may use it, and whose data and action permissions it uses.

Compare metered usage, not just subscriptions

A ChatGPT subscription does not imply unlimited Workspace Agent runs. For standard Business seats, these agents share an allowance and credit pool with other agentic features. Flexible pricing can provide additional usage through purchased credits; the allowance and agreement determine the extra bill.

Illustrative calculation, using rates checked November 15, 2026: suppose 10 people each run a workflow five times per workday for 20 days. That is 1,000 runs. Assume every run totals 50,000 uncached input tokens and 5,000 output tokens across all model calls. Tokens are the units of information a model processes; these quantities are assumptions, not measured consumption.

  • Workspace credits: the Workspace Agent row labelled GPT-5.6 lists 100 credits per million input tokens and 500 per million output tokens. Each assumed run uses (0.05 × 100) + (0.005 × 500) = 7.5 credits: 7,500 credits for the month, before included usage.

  • API model charges: the GPT-5.6 Sol Standard short-context rates are $4 per million input tokens and $20 per million output tokens. The assumed monthly totals are 50 million input and 5 million output tokens: (50 × $4) + (5 × $20) = $300 in model-token charges.

Some Enterprise customers use USD token-based workspace billing instead of credits. Its Workspace Agent row, labelled GPT-5.6 Sol, lists the same $4 input and $20 output base rates. The customer’s agreement governs the actual charge.

These figures do not establish which route is cheaper: a real migration can change models, token consumption and retries. The example excludes subscriptions, tools, storage, hosting, taxes, processing surcharges and migration work. It does not assign a universal dollar value to a workspace credit.

Recheck prices before committing a post-shutdown budget. The API pricing page describes GPT-5.6 Sol promotional pricing only through November 30; the displayed rates are not a December price promise.

Verify the replacement before switching

Use the old workflow’s intended results as the acceptance criteria. Suggested checks before sending live work to the replacement:

  • Run an ordinary request and each important branch. Check the resulting business action, not just the final message.

  • Try a user without permission, an action awaiting approval, a failed tool and a repeated request. Confirm that access and duplicate-action handling match your requirements.

  • Check scheduled runs, account connections and retained records. Assign an owner for failures and usage limits.

Also check dependencies: prompt objects and the Evals platform have their own November 30 retirements. Moving a workflow does not automatically resolve either. The separate guide to moving OpenAI prompt objects into code covers that prompt-storage task.

Finish and verify the replacement before November 30. The published shutdown notice does not specify an hour or timezone, and the migration guide does not establish export access after closure.