OpenAI Responses API: Migrating After the Assistants Shutdown

By Greg Nowak. Updated October 24, 2026.

If your support assistant or internal document workflow still depends on OpenAI’s Assistants API, the migration deadline has passed. OpenAI confirms that the API shut down on August 26, 2026, and is no longer available. Its Assistants migration guide now directs teams to Responses.

For a business owner, the immediate concern is whether customers and staff can finish their work. For the delivery team, the job is to restore that service while making clear decisions about history, tools, permissions, and ongoing cost. Start with one essential workflow and an agreed definition of success.

First, establish what still works

Inventory every integration that uses Assistants: website features, agency-built automations, scheduled jobs, internal scripts, and third-party connectors. Record the business owner, developer, affected users, and fallback for each. If a connector hides the implementation, ask its supplier which API it uses and whether the replacement is live.

Recover instructions, tool definitions, file references, and representative conversations from systems you control. OpenAI’s guide explicitly warns that the old thread-message retrieval call no longer works. Do not build a recovery plan around exporting history through retired endpoints.

If a workflow is already down, provide a visible manual route, such as a support queue or staff handoff. Make it clear who will handle outstanding requests.

Avoid the prompt-object migration detour

Parts of the Assistants guide still describe converting assistant configurations into reusable prompt objects. However, OpenAI’s prompt-object migration guidance says those objects are also deprecated, with v1/prompts scheduled to shut down on November 30, 2026.

For new production work, keep prompt content in version-controlled application code. Pass the generated instructions and input directly to Responses, and validate dynamic inputs. If your replacement already uses a saved prompt ID, include its removal in the plan.

Give prompt changes an owner and a release process. A change to refund guidance or document interpretation deserves review, testing, and a way to restore the previous version.

Choose where conversation history belongs

Use the simplest state model that meets the workflow’s needs. OpenAI’s conversation-state documentation describes the main options:

Response objects are saved for 30 days by default; store: false disables that Response storage. Conversation objects and their Items do not share the 30-day expiry. Document retention explicitly, including your own database and logs. An API setting alone is not a complete retention policy.

Rebuild one complete user journey

The useful conceptual map is Threads to Conversations, Runs to Responses, and Run Steps to typed Items. The application still needs to manage execution correctly. Preserve these implementation details from OpenAI’s Responses migration guide:

• Read typed output. Use the SDK’s output_text helper for final text; inspect response.output when handling tools or other Item types.
• Complete function-call loops. Return each function result with the matching call_id and preserve required context.
• Update schemas and streaming. Structured Outputs use text.format; streaming consumers must handle Responses event types.
• Resend stable instructions when chaining. previous_response_id does not carry forward the previous response’s top-level instructions.

Test the full journey from user request to business outcome. For example, a document assistant should find the permitted document, support its answer with evidence, and handle missing information sensibly.

Agree on acceptance checks before release

Build a small test set from approved examples. Include ordinary requests, ambiguous questions, missing documents, unauthorized access attempts, and unavailable tools. Have the operations owner review outcomes alongside the developer.

For retrieval, check document access and citations. For tools that change records or send messages, check authorization, approval where needed, and duplicate prevention during retries. Set timeouts and execution limits, and make failures visible to someone who can resolve them.

Release to a limited audience first. Compare correctness, completion time, errors, and cost per completed task. Since the old API is unavailable, the fallback must be a working replacement version or a manual process. Assign someone to activate it.

Budget for the whole workflow

As checked on October 24, 2026, OpenAI’s API pricing lists standard web search at $10 per 1,000 calls plus search-content tokens. File search costs $2.50 per 1,000 calls, with storage at $0.10 per GB per day after the first free GB. Model tokens are additional.

Container billing also deserves attention: eligible Hosted Shell and Code Interpreter sessions are billed by the minute with a five-minute minimum. Avoid budgeting every execution as a flat session charge.

Track tokens, tool calls, retries, and storage by workflow. Set application limits and alerts, then investigate expensive cases. Repeated searches and unnecessary conversation history can outweigh savings from choosing a cheaper model.

Leave the team with a service it can maintain

A useful handover includes versioned prompts, documented state decisions, acceptance tests, cost visibility, and a named support owner. Agency teams should also make account ownership and ongoing maintenance responsibilities explicit.

If you need help recovering a legacy assistant or getting a Responses replacement ready for production, Greg can help assess dependencies, coordinate implementation, and manage the cutover. Discuss your migration with Greg; bring the workflow, what is failing, and who currently maintains it.

Related on GrN.dk

• AI automations need a spend dashboard before the first runaway bill
• Not Every AI Job Needs an Instant Answer: Batch the Backlog
• OpenAI’s Assistants API Is Gone: What to Fix First

Need help with this kind of work?

Discuss your migration with Greg Get in touch with Greg.

Sources

• OpenAI Assistants migration guide
• Migrate from prompt objects
• Conversation state
• Migrate to the Responses API
• OpenAI API pricing