Guides · ← All guides · Model Retirement Calendar →

The October 2026 OpenAI Model Shutdown: Full List + Migration Map

Last updated: October 13, 2026 · 10 min read

By the Ingenious AI team

On October 23, 2026, OpenAI shuts down a dozen legacy model snapshots — the largest single retirement wave it has ever run. Announced on April 22, 2026 on its official deprecations page, the wave covers gpt-3.5-turbo, gpt-4, gpt-4-turbo, the o1 and mini reasoning line, and gpt-image-1. If your code was copied from a 2023–2024 tutorial, course, or Stack Overflow answer, there is a good chance it hard-codes one of these IDs.

First, the vocabulary, because OpenAI's own definitions carry the whole meaning. From the deprecations page: "When we announce that a model or endpoint is being deprecated, it immediately becomes deprecated. All deprecated models and endpoints will also have a shut down date. At the time of the shut down, the model or endpoint will no longer be accessible." Every model below has been deprecated since April. October 23 is when it is gone. And on shutdown day, OpenAI's rule is absolute: at the time of the shut down, the model or endpoint will no longer be accessible.

The full list

Twelve legacy snapshots plus five fine-tuned variants, straight from OpenAI's deprecations page (verified October 12, 2026). Note the second column: the deprecations page lists the bare aliases alongside the dated snapshots, so code calling gpt-4 rather than gpt-4-0613 is affected too. The alias does not save you on this one.

Fine-tuned variants retiring the same day: ft-gpt-3.5-turbo → gpt-5.6-terra, ft-gpt-4 → gpt-5.6-sol, ft-gpt-4.1-nano-2025-04-14 → gpt-5.6-luna, ft-babbage-002 → gpt-5.6-terra, ft-davinci-002 → gpt-5.6-terra. OpenAI's fine-tuning rule is explicit: inference on fine-tuned models keeps working only until the base model is deprecated — so every fine-tune on these bases has to be rebuilt, not just repointed.

Two entries that deserve a second look

Scanning the list for "GPT-4 and the o-series" misses two rows that bite differently:

• gpt-4o-2024-05-13 is a GPT-4o snapshot, not a GPT-4 one. If you pinned that snapshot in 2024 for reproducibility — which was the responsible thing to do — you are on this list. The teams that followed best practice are precisely the teams on the shutdown calendar.
• gpt-4.1-nano, gpt-4-turbo, and gpt-image-1 are listed as bare aliases rather than dated snapshots. Everything else is snapshot-pinned. And the alias column means the reverse trap: teams that deliberately used the alias to avoid pinning maintenance still retire on October 23, because OpenAI is retiring the alias names too.

The migration map: pick by workload, not by name

OpenAI names substitutes, but the right choice depends on what the workload actually does:

• General chat / classification / extraction (gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4o-2024-05-13) → gpt-5.6-sol or gpt-5.6-terra. Start with terra for cost; promote to sol where your evals show a quality gap.
• Reasoning workloads (o1, o1-pro, o3-mini) → gpt-5.6-sol (o1-pro users: set reasoning.mode: pro). Watch internal token burn — reasoning models bill the tokens you don't see.
• High-volume cheap calls (o4-mini, gpt-4.1-nano) → gpt-5.6-terra or gpt-5.6-luna. This is where the unit-economics check matters most — run a week of real traffic against the substitute and project the bill before the deadline.
• Image generation (gpt-image-1) → gpt-image-2.5-sunburst or gpt-image-2.5-flare. Regenerate your golden outputs — pixel-level drift is expected, and "close enough" is a decision you want to make deliberately.
• Fine-tunes → rebuild on the substitute base. The old weights are not portable across architectures; budget retraining or prompt re-engineering, not a config line. And start now: retraining has a lead time that a config change doesn't.

The trap most teams fall into

The model names your code contains are the easy part. The dangerous ones are the IDs your framework sends for you: LangChain and LlamaIndex have long defaulted to gpt-3.5-turbo when no model is specified — code that never mentions the model at all still breaks on October 23. Search your dependency configs and .env files, not just your source tree.

The second trap is the dashboard. If your prompts live in OpenAI's dashboard rather than your code, check for pmpt_… IDs or prompt names in your repo — the model choice is stored where your grep can't reach. And note the calendar: OpenAI's v1/prompts API shuts down November 30, 2026. Copy that content into your code regardless of which model wave you're handling.

What to do this week (before October 23)

Ten days out, triage beats perfection. In order:

1. Find every affected ID today. Run the 20-minute codebase audit — grep for the snapshot IDs and the bare aliases, check .env and CI secrets, cross-reference your OpenAI usage page grouped by model for the last 30 days. Anything in usage that wasn't in the grep is a shadow integration.
2. Classify by blast radius. Revenue-critical features get the full eval; internal tools get the straight substitute swap. Not every migration deserves the same ceremony.
3. Benchmark the substitute on your real prompts. Same prompt, different personality — run your top 20–50 production prompts against the replacement and diff the outputs before traffic moves. A name swap is not a migration.
4. Ship behind a flag with a rollback. Even this close to the date, a canary at 10% for 48 hours beats a midnight flip. Keep the old model ID in config until the new one is stable — the endpoint works until October 23, so use the remaining window.
5. Handle fine-tunes first, not last. They have the longest lead time (retraining) and the hardest deadline (the base model dies with them). If a fine-tune can't be retrained in time, the fallback is a well-engineered prompt on the substitute base — decide this week.

After October 23: the next waves

This wave is the start of the season, not the end. Already on the calendar: December 1, 2026 (gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest), December 11, 2026 (first-gen gpt-5 snapshots, o3, o3-pro), and November 30, 2026 for the v1/prompts API, the Evals platform, and Agent Builder — product sunsets, not model string swaps. If you migrate this month, pick a landing spot with its own runway: moving to a model that retires in December is doing the work twice. The full list lives on the Model Retirement Calendar.

FAQ

Which OpenAI models shut down on October 23, 2026?

Twelve legacy snapshots — gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-1106-preview, gpt-4-turbo, gpt-4.1-nano, gpt-4o-2024-05-13, gpt-image-1, o1-2024-12-17, o1-pro-2025-03-19, o3-mini-2025-01-31, o4-mini-2025-04-16, ft-o4-mini-2025-04-16 — plus five fine-tuned variants. Announced April 22, 2026 on OpenAI's deprecations page.

What happens if I still call a retired OpenAI model?

At the shutdown date the model or endpoint is no longer accessible — requests stop working, typically with a 404 and a model_not_found error body. There is no automatic substitution.

Do my fine-tuned models survive the base model's retirement?

No. Inference on fine-tuned models is available only until the base model is deprecated, so every fine-tune built on a retiring base must be rebuilt on the substitute base.

Which replacement should I pick?

OpenAI's mapping: gpt-5.6-sol for the GPT-4 family, o1, o3-mini and vision work; gpt-5.6-terra for o4-mini, GPT-3.5 and completions-era workloads; gpt-5.6-luna for the nano tier; gpt-image-2.5-sunburst/flare for images. Always benchmark the substitute on your own prompts before switching — a name swap is not a migration.

Can I rely on the bare alias instead of the snapshot?

Not for this wave. The deprecations page lists the bare aliases (gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o1-pro, o3-mini, o4-mini, gpt-4.1-nano) with the same October 23 shutdown date. The alias retires with the snapshot.

Never scramble through one of these again

ModelMinder watches the deprecation feeds so you don't have to: track the models your product depends on, get alerted before the shutdown date, and run a one-click migration eval — your prompts against the replacement, diffed and judged — before you move production traffic.

If you're reading this close to the deadline

Shutdown dates have a way of arriving faster than migration plans. If the retirement is weeks away and you haven't started, triage beats perfection. Rank your model references by blast radius: user-facing features first, internal tooling last. Migrate the top of the list with full evals; the bottom of the list can move with lighter validation.

Don't skip the eval on the critical path, even under time pressure. A rushed swap that changes your product's behavior is worse than a planned migration that finishes a week late — and providers sometimes offer short extensions for enterprise workloads, though you should never count on one. The eval is also your evidence if something goes wrong: "we tested the replacement on our real prompts and it passed" is a very different incident review than "we swapped the ID and hoped."

After the cutover, keep the old model ID in your codebase search for one more cycle. Stragglers hide in cron jobs, notebooks, and third-party integrations that pin the model on your behalf. The shutdown only hurts once if you verify it's actually gone.