OpenAI's Oct 23 Model Shutdown: Your Migration Checklist

TL;DR: OpenAI shuts down 11 legacy model snapshots plus all fine-tuned variants this Friday, 23 October 2026 — announced 22 April 2026 on OpenAI's official deprecations page. Every GPT-4 and GPT-3.5 name, the original gpt-4o-2024-05-13 snapshot, o1, o1-pro, o3-mini, o4-mini, gpt-4.1-nano and gpt-image-1 stop answering; named substitutes are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna (images move to the gpt-image-2.5 line). The short aliases — gpt-4, gpt-4-turbo, gpt-3.5-turbo, o1, o3-mini, o4-mini — die too.

Two days. That's the entire migration window left. If any of the strings below appear in your code, your environment variables, your batch files, or your scheduled jobs, requests to them fail on Friday. Not degrade — fail. This is the migration checklist version of our running Deprecation Watch calendar: same October 23 deadline, but every ID, every alias, every substitute, and the exact steps to be off the old names before the shutdown.

Everything shutting down this Friday#

The table below is copied from OpenAI's deprecations page, section "2026-04-22: Legacy GPT model snapshots", re-read and verified on 21 October 2026. Read it by the exact string your code sends — each row's aliases go down with it.

All shutdown dates verified 21 October 2026 against OpenAI's deprecations page.

Three things about this table catch people out:

1. The most dangerous row is gpt-4o-2024-05-13 — and only that one. The other GPT-4o snapshots (gpt-4o-2024-08-06, gpt-4o-2024-11-20) and gpt-4o-mini-2024-07-18 are not on Friday's list; they retire 14 January 2027 (same page, verified 21 October 2026). Don't migrate the wrong snapshot early — and don't assume Friday's done deal covers the rest of the family.
2. gpt-image-1 users have the longest migration. The substitute isn't a text model at all — it's the gpt-image-2.5 line, which means regenerating image pipelines, not swapping a model string.
3. The alias column is the point. Codebases that migrated "off GPT-4" years ago often still pass gpt-4 — the short name — in a wrapper default. It's an alias for the retiring gpt-4-0613, and it stops answering Friday.

Your fine-tunes go the same day#

Fine-tuned versions of the retiring families are removed on the same shutdown date. OpenAI lists these explicitly (same page, verified 21 October 2026):

There is no automatic conversion: a fine-tune is a separate model object, and it dies with its family. If production traffic runs through a fine-tune, the work isn't a model-string swap — it's evaluating the substitute base (or a new fine-tune of it) on your eval set before Friday.

What actually happens at shutdown#

Requests to a retired model fail. There is no grace mode and no degraded fallback — a retirement you miss is a production outage, full stop. OpenAI's page says access "will be shut down" on the listed dates without describing the exact error shape (as the independent tracker BenchLM notes, we haven't seen a tested report of the specific error code). Plan for a hard failure — a 4xx/5xx on every request naming the old ID — and treat anything softer as luck.

One more platform caveat, also from BenchLM's cross-check of the two providers' pages (read 13 October 2026): the date that matters is the one on the platform you actually call. On Azure AI Foundry the same snapshots retire on different dates — gpt-4o-2024-05-13 runs to 9 December 2026 there, and the o1/o3-mini versions to 19 November 2026. Microsoft's retirement page states that retired Azure deployments return 410 Gone and that dates are not extendable. If you call Azure, Friday is not your deadline — but don't read that as extra time to waste: verify the date in your own deployment's model page.

The 4-step migration checklist#

Do these in order. Today is Wednesday; the shutdown is Friday.

1. Inventory every retiring ID in your code — today. Model strings hide in more places than the chat setting: environment variables, shared SDK wrappers with default arguments, scheduled jobs and queue workers that deploy separately, batch input files (each line names its model), eval notebooks, and gateway fallback configs. This finds the strings, and also catches gpt-4o-mini and the later GPT-4o snapshots so you can classify rather than assume:

rg -n -e 'gpt-3\.5-turbo' -e 'gpt-4(-0613|-1106-preview|-turbo)?' -e 'gpt-4\.1-nano' \
  -e 'gpt-4o-2024-05-13' -e 'gpt-image-1' -e '\bo1(-pro)?\b' -e 'o3-mini' \
  -e 'o4-mini' -e 'ft-' --glob '!node_modules' .

2. Map each hit to its named substitute and note the breaking differences. The swap is not always one line:

• o1-pro → gpt-5.6-sol needs reasoning.mode: pro to match behavior.
• The 5.6 family has explicit reasoning-effort settings (none / low / medium / high / xhigh / max); a comparison that inherits defaults changes two things at once, so pin effort when testing.
• Substitutes have different context windows and rate limits — check both before Friday, not after.
• gpt-image-1 → gpt-image-2.5-sunburst/flare is a pipeline rebuild, not a string swap.

3. Test the substitute on your own workload, then move traffic in stages. Freeze the prompt, tools, and output schema; replay real inputs (50–100 per job, secrets stripped) and set pass/fail checks before you run: output parses, validates against the schema, and field accuracy stays within your tolerance of the current model. Move traffic shadow → canary (1–5%) → ramp → full, with a written rollback trigger (e.g. three consecutive failures, or p95 latency over your limit). Note the fine print BenchLM flags: Sol's current rate is promotional — committed only through 21 November 2026 — so re-run your cost arithmetic after that date.

4. Set a per-key spend cap before the switch. A model swap always moves the bill — sometimes down, sometimes up — and the first month on a new model is when a misconfigured agent burns budget. Put the migrating workload on its own key with a monthly cap so one bug can't spend the whole budget (see spend caps for coding agents).

What the substitutes cost on tokens.bd#

You don't need a US card to run OpenAI's named substitutes. On the tokens.bd catalog today — checked 21 October 2026 — two of the three named substitutes are live under stable IDs:

| Substitute | Input / 1M | Output / 1M | Replaces (retiring Friday) | | --- | --- | --- | | openai/gpt-5.6-sol | $5.50 (৳679.80) | $33.00 (৳4,078.80) | gpt-4 family, gpt-4o-2024-05-13, o1, o1-pro, o3-mini | | openai/gpt-5.6-luna | $0.22 (৳27.19) | $1.32 (৳163.15) | gpt-4.1-nano |

Prices checked 21 October 2026 on tokens.bd/models; BDT at ৳123.60 per USD (mid-market, checked 21 October 2026). BDT arithmetic: $5.50 × 123.60 = ৳679.80; $33.00 × 123.60 = ৳4,078.80; $0.22 × 123.60 = ৳27.19; $1.32 × 123.60 = ৳163.15.

Two honest caveats, because a migration guide that hides them is a trap:

• gpt-5.6-terra — OpenAI's named substitute for gpt-3.5-turbo, o4-mini, and the ft-gpt-3.5-turbo / ft-babbage-002 / ft-davinci-002 fine-tunes — is not on tokens.bd's catalog today. If your workload runs on one of those, you'll need OpenAI access for the exact substitute, or pick a different cheap model from the catalog and evaluate it against your workload.
• The image substitutes (gpt-image-2.5-sunburst, gpt-image-2.5-flare) are not on tokens.bd either — the catalog carries no image models. gpt-image-1 callers are the ones with the real rebuild ahead of them.

One more thing worth knowing about the swap: switching models resets your prompt cache — caches are per model, so the new substitute pays full input price on the whole context until the cache warms up. Budget a one-time full-price period into the first days after the cutover (see prompt caching explained).

Don't stop at Friday — three more waves are already scheduled#

Friday is the first of four OpenAI waves already on the official calendar (all re-verified against the deprecations page on 21 October 2026):

• 1 December 2026: gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest → the gpt-image-2.5 line.
• 11 December 2026: the first-gen GPT-5 snapshots (gpt-5-2025-08-07, gpt-5-mini, gpt-5-nano, gpt-5-pro-2025-10-06) and o3 / o3-pro → gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna (per model).
• 14 January 2027: the rest of the GPT-4o and GPT-4.1 family (gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14, gpt-4.1-mini) → gpt-6-luna.

The full running calendar — including Anthropic, Google, and the platform shutdowns — lives in AI Deprecation Watch: November 2026 Deadline Calendar, which links back to this guide for the October 23 deep dive.

The same wider deprecation wave also reaches evaluation tooling: OpenAI Evals goes read-only on 31 Oct, so work through the OpenAI Evals migration checklist before that deadline as well.

October 23 shutdown FAQ#

What exactly happens to my requests on Friday — do they fail or just get worse? They fail. There is no grace mode and no degraded fallback. OpenAI's page says access "will be shut down" on the listed dates without documenting the exact error shape, so plan for a hard failure on every request naming an old ID. A retirement you miss is a production outage — which is why the inventory grep above matters more than any migration guide.

Do the short aliases like gpt-4 and gpt-3.5-turbo shut down too? Yes. They're aliases for retiring snapshots (gpt-4 → gpt-4-0613; gpt-3.5-turbo → gpt-3.5-turbo-0125), and OpenAI lists them in the shutdown rows. This is the trap that catches codebases that thought they migrated years ago — a wrapper default still passing gpt-4 stops answering Friday.

I'm on Azure — is Friday my deadline? Not necessarily. Azure AI Foundry sets its own dates: gpt-4o-2024-05-13 runs to 9 December 2026 there, and the o1/o3-mini versions to 19 November 2026 (per Microsoft's schedule, cross-checked by BenchLM on 13 October 2026). Microsoft also notes retired deployments return 410 Gone and dates aren't extendable. Read the date on the platform you actually call.

Can I keep using my fine-tuned gpt-4 after Friday? No. Fine-tuned variants (ft-gpt-4, ft-gpt-3.5-turbo, ft-gpt-4.1-nano-2025-04-14, ft-babbage-002, ft-davinci-002, ft-o4-mini-2025-04-16) are removed the same day. There's no automatic conversion — evaluate the recommended replacement base (or a fresh fine-tune of it) on your eval set before the shutdown.

Where can I run the substitutes without a US card? gpt-5.6-sol and gpt-5.6-luna are live on tokens.bd's 60+ model catalog under stable IDs, payable in BDT via bKash/Nagad. Two gaps to know: gpt-5.6-terra (the named substitute for gpt-3.5-turbo and o4-mini) and the gpt-image-2.5 substitutes aren't on the catalog today.

All shutdown dates and substitutes verified 21 October 2026 against OpenAI's deprecations page; catalog prices checked the same day on tokens.bd/models. Providers change schedules — the deprecations page is the authority.

Sources: OpenAI deprecations page · BenchLM: GPT-4o Retirement Date — OpenAI's October 23 API Shutdown · tokens.bd model catalog

Once your model IDs survive Friday, here's how to decide where to buy your OpenAI API access: OpenAI on Google Cloud: The Developer's Buying Guide

As a next step after the shutdown wave, consider GPT-6.1 Terra: $1 Computer Use, Explained as the current mid-tier OpenAI option for computer-use work.

Google's turn came next — see Gemini 2.5 Deprecated: Your Dec 1 Migration Checklist.

See also: Kimi K3.1 Pro Open Weights Are Live: 1.56 TB Guide.

See also: our explainer on OpenAI's newest gated model, GPT-6 Cyber: Price, Access, and What It Can Do.