OpenAI Pulls the Plug on 29 Legacy Models on October 23: The n8n Migration Checklist

On October 23, 2026 — 14 days from now — OpenAI will switch off 29 legacy model IDs and snapshots in one go. That includes every GPT-4 and GPT-3.5 Turbo name your workflows probably still call: gpt-3.5-turbo, gpt-4, gpt-4-turbo, the o-series reasoning models (o1, o1-pro, o3-mini, o4-mini), and gpt-4.1-nano. There is no graceful degrade: at shutdown time, OpenAI says, the model “will no longer be accessible” — your pinned calls fail outright with an API error.

If your n8n instance has an OpenAI node with the model field set to gpt-4-0613 or gpt-3.5-turbo-0125, that workflow breaks on the 23rd. This guide walks you through a full audit-and-migrate pass: find every dead model string in your workflows, swap in the correct substitute, and verify before the deadline.

By the end, your n8n instance will have zero references to the October 23 shutdown list, and a test run will prove each changed workflow still works.

What dies on October 23

OpenAI announced this wave on April 22, 2026, under the heading “Legacy GPT model snapshots.” Every entry below has a shutdown date of 2026-10-23 (verified on OpenAI’s deprecations page, via DigitalMatters and the modeldeprecations.dev mirror):

Three smaller waves already passed — September 24 (the entire Sora 2 family and Videos API, with no replacement offered), September 28 (the last GPT-3-era names), October 1 (gpt-5.4-cyber). OpenAI has shown it follows through. This wave has been “deprecated” since April; October 23 is when the models are simply gone.

Prerequisites

• A self-hosted or cloud n8n instance with workflows that call OpenAI (the OpenAI node, the AI Agent node with an OpenAI model, or HTTP Request nodes hitting the API directly).
• An OpenAI API key with access to the GPT-5.6 family substitutes listed above.
• jq installed if you want the fast audit route, and the ability to export workflows from n8n.

Example terminology — replace these placeholders in every command:

• <dead-models> — one of the model strings from the table above.
• <n8n-host> — your n8n host, e.g. https://n8n.example.com.
• <api-key> — your OpenAI API key (keep it out of chat logs and version control).

Step 1 - Audit every workflow for dead model strings

The model you named in your code is easy to find. What your frameworks chose for you is not: LangChain’s ChatOpenAI() with no model= argument still sends gpt-3.5-turbo by default — a request that dies on October 23 even though you never typed the name.

The fastest reliable audit is a text search over your exported workflow JSON. Export each workflow from the n8n UI (Workflow menu → Export → Download), then run:

grep -rEo '"model":\s*"[^"]*"' workflows/ | sort | uniq -c | sort -rn

This prints every model string your workflows reference, with counts — scan the output against the shutdown table above. In the n8n UI, you can also check each AI Agent node’s model selector, but the grep catches what the UI misses: pinned model strings inside credentials, Code nodes, and HTTP Request bodies.

Note: Bare aliases count. If your code says gpt-4.1-nano or o4-mini rather than the dated snapshot, it is still on this list — OpenAI listed those names as their own rows. And gpt-4o-2024-05-13 is a GPT-4o snapshot, not GPT-4: if you pinned it in 2024 for reproducibility, you are affected.

You should see either zero matches against the shutdown list (you are already safe) or a list of files to fix. Keep that list — it becomes your migration checklist.

Step 2 - Map each dead model to its substitute

Use the mapping table above, choosing the OpenAI-recommended substitute per row. Two decisions to make consciously:

• Cost. The 5.6 family has different pricing than the 2023-era models it replaces — a workflow running millions of gpt-4-0613 calls per month will not cost the same on gpt-5.6-sol. Check OpenAI’s current pricing page before migration day and budget for it, rather than discovering the difference on the 24th.
• Quality. gpt-5.6-terra is OpenAI’s suggested replacement for the cheap tier (3.5-turbo, o4-mini), not a clone of it. Run your changed workflows with real inputs (Step 4) before you trust the swap in production.

Also decide now whether to test gpt-6 Sol/Luna as well: OpenAI’s April guidance pointed to GPT-5.6, but for the newest deprecation waves it points to GPT-6, and GPT-6 pricing was cut at launch. If you are rewriting prompts anyway, one comparison run is cheap.

Step 3 - Swap the models in n8n

Open each flagged workflow and change the model in the OpenAI node’s Model field (or the equivalent parameter in a Code/HTTP node). In n8n:

1. Open the workflow, click the AI Agent or OpenAI node.
2. Change Model to the substitute from the table.
3. If the node used a model pinned in an expression (e.g. ={{ $json.model }}), trace where that value comes from and update it at the source.
4. Save and activate.

Warning: Do this even for workflows that are currently turned off. OpenAI applies shutdown to the model, not your usage pattern — a disabled workflow you re-enable on October 24 fails silently at 3 AM. The grep in Step 1 already found them; don’t skip them now.

If you use LangChain inside a Code node, add the explicit model= parameter rather than relying on the default — the default was the trap to begin with.

Step 4 - Rebuild any fine-tuned models now — they cannot be converted

This is the one step with no shortcut. Fine-tuned variants (ft-gpt-3.5-turbo, ft-gpt-4, ft-gpt-4.1-nano-2025-04-14, ft-o4-mini-2025-04-16, ft-babbage-002, ft-davinci-002) shut down alongside their base models, and OpenAI states that inference on fine-tuned models keeps working only until the base model is deprecated. You must retrain on a current base (gpt-5.6-terra, gpt-5.6-sol, or gpt-5.6-luna) — there is no export-and-import path.

If retraining is impossible before October 23, prepare a fallback: point the workflow at the base substitute with few-shot examples in the prompt, and schedule the retrain as debt, not a plan.

Step 5 - Verify every changed workflow

For each migrated workflow, run one test execution with realistic input and check the result — not just that it completes, but that the output quality matches what the old model produced. Then:

1. Confirm the OpenAI Usage dashboard shows calls hitting the new model names.
2. Set a one-time calendar reminder for October 23 to check error logs — you want to find any missed reference that day, not when a customer does.
3. Pin the new model name explicitly everywhere. You have now learned why defaults and bare aliases are a liability.

Hard cases and limits

• ChatGPT is not affected by this shutdown — it applies to the API, not the consumer product. Azure OpenAI has its own retirement timeline; check your Azure portal if you use it.
• I did not verify substitute pricing or quality myself. The substitutes are OpenAI’s official recommendations, not tested benchmarks — measure before you trust them at scale.
• The gpt-image-1 replacement is the vaguest row: OpenAI lists gpt-image-2.5-sunburst or -flare in one source and gpt-image-2 in another. Test image generation outputs before the 23rd if visuals are part of your product.
• This post covers the October 23 wave only. December brings the shutdown of the gpt-5 August snapshots (gpt-5-2025-08-07, mini/nano/pro, o3-2025-04-16) — keep the grep from Step 1; you will run it again.

FAQ

What happens to a workflow still calling gpt-4 on October 23?

The API call fails — OpenAI removes the model entirely, so there is no graceful degradation and no automatic redirect to a newer model. Every affected n8n workflow errors on its next run.

Is the bare name gpt-4 affected, or only dated snapshots like gpt-4-0613?

Both. OpenAI listed short names and bare aliases (gpt-4, gpt-4-turbo, gpt-4.1-nano, o1, o1-pro, o3-mini, o4-mini, gpt-3.5-turbo) as their own rows with the same October 23 shutdown date.

Do I need to migrate if my workflows only use GPT-5 models?

No. This wave touches legacy GPT-3.5/4, the o-series snapshots, gpt-4.1-nano, and gpt-image-1. Current GPT-5.6 and GPT-6 models are the safe targets.

What about my fine-tuned gpt-3.5-turbo model?

It shuts down on October 23 with the base model. You must retrain on a supported base (gpt-5.6-terra, -sol, or -luna); there is no conversion path. Start now — retraining is the longest lead-time item on this checklist.

Key Takeaways

1. 29 model IDs and snapshots go dark on October 23 — gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4.1-nano, the o-series, and all their fine-tunes, plus gpt-image-1.
2. Grep your exported workflow JSON first — the audit takes minutes and catches the traps: bare aliases, framework defaults, and pinned snapshots.
3. Swap to OpenAI’s listed substitutes (mostly gpt-5.6-sol, terra for the cheap tier, luna for nano) and test output quality before trusting them.
4. Fine-tunes must be rebuilt — there is no migration path, only retraining. This is your longest lead time; start it today.

Next: keep the audit grep in a runbook and re-run it quarterly — the next wave (December, the gpt-5 August snapshots) is already scheduled. For the bigger picture on keeping self-hosted n8n healthy, read our n8n October 2026 Security Update.

Images: Pexels