OpenAI's October 23, 2026 Model Shutdown: What Goes Dark and How to Migrate

In April 2026, OpenAI emailed developers a deprecation notice that read like a rout: the GPT-3.5 and GPT-4 lines, the original GPT-4o, the entire first generation of o-series reasoning models, and the gpt-image-1 API — all scheduled to lose API access on October 23, 2026. That date is now eleven days away. If any of your production code still names one of these models, it will start returning errors that morning. Here's the confirmed list, the migration targets, and a checklist you can actually finish in time.

Most OpenAI deprecations are quiet: a model stops appearing in the picker, a docs page gains a strikethrough. The April 2026 notice was different in scale. It retired, in one list: gpt-3.5-turbo, gpt-4-0613, gpt-4-turbo, gpt-4.1-nano, gpt-4o-2024-05-13 (the original GPT-4o), gpt-image-1, o1-2024-12-17, o1-pro, o3-mini, and o4-mini — with API access ending October 23, 2026. Developer forum threads described it as an "unprecedented massive list", and the reason is simple: OpenAI is not retiring failures. It's retiring the models an entire generation of production code was built on.

The October 23 shutdown list

These models lose API access on October 23, 2026, per OpenAI's notice:

Migration targets follow OpenAI's stated pattern: each deprecated model's recommended replacement is its closest current-generation descendant. Always confirm the exact target on OpenAI's live deprecations page before editing code.

The fine-tuned versions of gpt-3.5-turbo, gpt-4, gpt-4.1-nano, babbage-002, davinci-002 and o4-mini lose access the same day. That last category matters more than it sounds: a fine-tune you trained in 2024 and have been calling quietly for two years dies with its base model. Rebuilding one is not a config edit — it's a training job against a new base model, with re-validation of outputs.

Why this shutdown wave is different

Three things distinguish this round from routine model rotation:

• It removes the "safe old fallback." For three years, gpt-4o-2024-05-13 was the model teams pinned when a newer model misbehaved. That escape hatch disappears on October 23 — incident runbooks that say "fall back to 4o" need a rewrite.
• It kills the entire first o-series. o1 and o1-pro were the proof that reasoning models were a product category. Their retirement means every reasoning workload must move to a successor — and successors handle prompts differently, especially around system messages and reasoning effort.
• It ends gpt-image-1. The first OpenAI image API model — the one behind a million "AI sticker" features — goes dark. And a second wave is already scheduled: gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest shut down December 1, 2026, consolidating everything on gpt-image-2.

A migration checklist that fits in an afternoon

For most small teams, the actual work is smaller than the scare. Here's the sequence:

• 1. Inventory (30 minutes). Search your codebase and infrastructure config for every deprecated model string: gpt-3.5-turbo, gpt-4-0613, gpt-4-turbo, gpt-4.1-nano, gpt-4o-2024-05-13, gpt-image-1, o1-2024-12-17, o1-pro, o3-mini, o4-mini-2025-04-16. Don't forget Terraform, Helm charts, n8n/Zapier workflow configs, and that one cron script nobody owns.
• 2. Map each call site to its target (30 minutes). Chat workloads move to the current mainline model; image calls move to gpt-image-2; reasoning calls move to the current o-series successor. Match on capability, not price alone.
• 3. Diff behavior, not just status codes (1–2 hours). Run your ten most common real prompts against the old and new models and compare. Reasoning-model successors may change response structure, token counts, or latency. If you parse specific output formats, this is where breaks hide.
• 4. Update and canary (1 hour). Change the model strings, deploy to a staging route or a small traffic slice first, watch error rates for a day if your timeline allows — the deadline is close enough that a same-day cutover may be necessary, but never all-at-once across every service.
• 5. Update the fallback runbook (15 minutes). Anywhere your incident docs name a deprecated model as the fallback, write in the new one. The next outage is a bad time to discover your safety net was retired in October.

What breaks if you do nothing

On October 23, API calls naming a shutdown model return an error — the request is rejected, not downgraded. In practice that means: chat features throw 4xx errors, image generation endpoints fail, scheduled jobs die silently (or loudly), and any third-party tool that hardcoded an old model string starts failing at its next scheduled run. The failure mode is total, which is why the fix is worth doing today rather than on October 22.

The pattern to watch

Zoom out and a strategy is visible: OpenAI has spent 2026 compressing its catalog — GPT-5.6 and GPT-6 tiers replaced the sprawling 4-series, one image API is replacing four, and legacy snapshots are being cleared on a published schedule. The lesson for anyone building on commercial AI APIs is not "avoid OpenAI". It's: never hardcode a dated model snapshot in a long-lived system. Pin a model family and a date, subscribe to the provider's deprecation feed, and keep a one-line config indirection between your product and any vendor's model name. That indirection costs one hour now and saves a migration every year.

One-line takeaway: October 23, 2026 retires the models your 2024 architecture was designed around — inventory today, canary tomorrow, and treat every vendor model string in production as a liability with an expiry date.

Prices and benchmark figures are standard published rates as of October 12, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.