Two retirement dates land this month, and they come from different companies with different rules. On October 19, GitHub removes six models from every Copilot surface, including GPT-5.5 and GPT-5.4. GitHub changelog, September 18 On October 23, OpenAI shuts off API access to gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4o-2024-05-13, gpt-4.1-nano, o1, o1-pro, o3-mini, o4-mini, gpt-image-1 and the fine-tunes built on them. OpenAI deprecations page Neither date is new. GitHub gave 31 days’ notice; OpenAI announced its list on April 22, six months out.
What is new is a pair of notices OpenAI posted on October 1: gpt-5.1, gpt-5.3-codex and gpt-5.4-nano shut down on April 1, 2027, and tts-1, tts-1-hd and both gpt-4o-mini-tts snapshots shut down on January 6, 2027. OpenAI deprecations page Anyone who moved to GPT-5.1 as a GPT-4 replacement now has a second migration queued.
This article is the checklist. We re-read the deprecations page, the pricing page and the GitHub post on October 7; every price below comes from OpenAI’s pricing page on that day, and S5 Labs has not tested any of the replacement models. OpenAI pricing
The October 23 list, with what OpenAI wants you to use instead
OpenAI names a substitute for each model. The substitutes are not always cheaper. The table pairs each retired model with OpenAI’s recommendation and the standard-tier rates for both, in dollars per million tokens, input then output.
| Retired on Oct 23 | Current rate (in / out) | OpenAI’s substitute | Substitute rate (in / out) |
|---|---|---|---|
gpt-3.5-turbo (alias for -0125) | $0.50 / $1.50 | gpt-5.6-terra | $2 / $12 |
o4-mini | $1.10 / $4.40 | gpt-5.6-terra | $2 / $12 |
gpt-4 (alias for -0613) | $30 / $60 | gpt-5.6-sol | $4 / $20 |
gpt-4-turbo | $10 / $30 | gpt-5.6-sol | $4 / $20 |
gpt-4o-2024-05-13 | $5 / $15 | gpt-5.6-sol | $4 / $20 |
o1 | $15 / $60 | gpt-5.6-sol | $4 / $20 |
o3-mini | $1.10 / $4.40 | gpt-5.6-sol | $4 / $20 |
o1-pro | $150 / $600 | gpt-5.6-sol with reasoning.mode: pro | not listed separately |
gpt-4.1-nano | $0.10 / $0.40 | gpt-5.6-luna | $0.20 / $1.20 |
gpt-4-1106-preview | not on the pricing page | gpt-5.6-sol | $4 / $20 |
gpt-image-1 | $10 image in / $40 image out | gpt-image-2.5-sunburst or -flare | $8 image in / $30 image out |
Sources: deprecations page for the pairings, pricing page for every rate.
Three things in that table deserve a closer look before you accept the default pairing.
Most GPT-4-era prices fall. A gpt-4 call at $30 and $60 moves to $4 and $20; o1 goes from $15 and $60 to the same $4 and $20. The exception is gpt-4o-2024-05-13, whose output rate rises from $15 to $20. Those Sol rates are promotional, though. OpenAI’s pricing page says GPT-5.6 Sol’s promotional pricing “is available at least through November 21, 2026”, and the model page describes the current rates as a 20% input and 33% output reduction from the earlier tier. GPT-5.6 Sol model page Budget on the possibility that Sol costs more after November 21 than it does on October 23.
The cheap-tier prices rise. gpt-3.5-turbo to gpt-5.6-terra is four times the input rate and eight times the output rate. o3-mini to gpt-5.6-sol is roughly 3.6 times on input and 4.5 times on output. gpt-4.1-nano to gpt-5.6-luna doubles input and triples output. Nothing obliges you to take the model OpenAI names. For a gpt-3.5-turbo workload that was chosen because it was cheap, gpt-5.6-luna at $0.20 and $1.20 is the obvious first thing to test, with Terra as the fallback if quality is not there.
Rate comparisons also ignore reasoning tokens. The Sol model page lists reasoning.effort values of none, low, medium (the default), high, xhigh and max, and reasoning output is billed as output. GPT-5.6 Sol model page A non-reasoning gpt-3.5-turbo prompt that returns 200 tokens may return many more under a default-effort reasoning model. Set the effort explicitly in the migration and compare bills on real traffic, not rate cards.
Two smaller notes. o1-pro’s substitute is Sol with reasoning.mode: pro; that parameter appears on the deprecations page, but the Sol model page we read today lists only reasoning.effort, and the pricing page has no pro-mode rate. Confirm the parameter against the API reference before you ship it. And gpt-4-1106-preview appears on the deprecations page twice: once in the March 26, 2026 shutdown list and again under October 23. Treat it as already gone.
Find every pinned ID, including the ones outside your code
OpenAI says it emails customers who are actively using the model. OpenAI deprecations page Whoever receives those emails is not always the person who will fix the code, so do your own search.
The model IDs are plain strings, which makes this a grep job:
cat > retiring-ids.txt <<'EOF'
gpt-3.5-turbo
gpt-3.5-turbo-0125
gpt-3.5-turbo-completions
gpt-4
gpt-4-0613
gpt-4-completions
gpt-4-0613-completions
gpt-4-turbo
gpt-4-turbo-2024-04-09
gpt-4-turbo-completions
gpt-4-1106-preview
gpt-4o-2024-05-13
gpt-4.1-nano
gpt-4.1-nano-2025-04-14
gpt-image-1
o1
o1-2024-12-17
o1-pro
o1-pro-2025-03-19
o3-mini
o3-mini-2025-01-31
o4-mini
o4-mini-2025-04-16
EOF
grep -rnwF -f retiring-ids.txt . --exclude-dir=node_modules --exclude-dir=.gitThe -w flag stops gpt-4 from matching gpt-4o, but it still matches gpt-4 inside gpt-4.1 and gpt-4.1-mini, and o1 inside o1-mini, because grep treats the dot and hyphen as word boundaries; read the hits rather than counting them. The ID list is OpenAI’s own alias table, so gpt-4 and gpt-4-0613 are the same model and both need to go. Plain-text aliases, deprecations page
The code is the easy part. The IDs that survive a migration tend to live elsewhere:
- Environment variables and secrets managers, where
OPENAI_MODEL=gpt-4-turbohas sat since 2024. - Workflow tools and CRM automations that call OpenAI through a built-in connector and store the model name in a dropdown.
- Spreadsheet add-ins, Zapier-style steps and internal admin pages with a hard-coded default.
- Batch jobs that were written once and run on a schedule, where the first sign of failure will be an empty output file on October 24.
- Fallback chains. A retry path that drops from a current model to
gpt-4on error will start failing on the retry.
Note what is not on the list: the gpt-4o alias, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini are not deprecated and remain on the pricing page. Only the May 2024 gpt-4o snapshot goes. OpenAI pricing The February retirement of GPT-4o from ChatGPT, which we covered in the Frontier Alliance piece, was a consumer-product change and left the API alone.
If a workload cannot move in time, OpenAI says some developers “may be able to provision dedicated capacity for continued access after a model’s shutdown date” through its sales team. OpenAI deprecations page That is a sales conversation rather than a setting, so start it this week if you need it.
Fine-tunes go with their base, and GPT-5.6 is not on the fine-tuning price list
Fine-tuned models die with their base. On October 23 OpenAI removes 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 and ft-davinci-002, and lists a “recommended replacement base model” for each: Terra for the 3.5-turbo, o4-mini, babbage and davinci fine-tunes, Sol for GPT-4, Luna for the nano. OpenAI deprecations page
Read that column carefully. None of the GPT-5.6 models appears in the fine-tuning price table; the models listed there are o4-mini-2025-04-16, the gpt-4.1 family, two gpt-4o snapshots and the legacy trio. Fine-tuning pricing So “replacement base model” means the model you prompt instead, not a model you retrain on your old dataset. If your fine-tune existed to teach a format or a classification scheme, the migration is a system prompt and a set of examples, then an evaluation against the fine-tune’s held-out set while it still answers.
The window for retraining anything is also closing. OpenAI calls this a wind-down: organizations that had never run fine-tuning lost job creation on May 7, organizations with no fine-tuned inference in the previous 60 days lost it on July 2, and active customers lose it on January 6, 2027. Inference continues until the base model is deprecated. Self-serve fine-tuning update A gpt-4.1-mini fine-tune is safe for now, because neither it nor its base is on any shutdown list, but you cannot count on creating a new one next year.
The October 1 notices: April 2027 and January 2027
The second migration arrives before the first one is cold. OpenAI’s stated minimums are six months’ notice for generally available models and three for specialized variants; the October 1 notices give the three text models six months and the speech models at least three. Model deprecation notice periods
| Retired | Date | Current rate (in / out) | Substitute | Substitute rate (in / out) |
|---|---|---|---|---|
gpt-5.1 | Apr 1, 2027 | $1.25 / $10 | gpt-6-sol | $2 / $10 |
gpt-5.3-codex | Apr 1, 2027 | $1.75 / $14 | gpt-6-sol | $2 / $10 |
gpt-5.4-nano | Apr 1, 2027 | $0.20 / $1.25 | gpt-6-luna | $0.10 / $0.50 |
tts-1 | Jan 6, 2027 | $15 per 1M characters | gpt-realtime-2.1-mini | $0.60 text in / $20 audio out |
tts-1-hd | Jan 6, 2027 | $30 per 1M characters | gpt-realtime-2.1-mini | $0.60 text in / $20 audio out |
gpt-4o-mini-tts (both snapshots) | Jan 6, 2027 | $0.60 text in / $12 audio out | gpt-realtime-2.1-mini | $0.60 text in / $20 audio out |
Sources: deprecations page, pricing page.
The text models are the straightforward half. GPT-6 Sol costs more per input token than GPT-5.1 and less per output token than GPT-5.3-Codex; our GPT-6 Sol and Luna article covers its million-token context and the cache behaviour that changes the sums for agent loops. The nano move is a price cut on both sides.
The text-to-speech half asks for more than a model swap. tts-1 is billed per character through the speech endpoint; gpt-realtime-2.1-mini is billed per token through the Realtime API, which OpenAI’s migration note points you to. Realtime API guide A batch job that turns articles into MP3 files will need rewriting around sessions, and the per-character and per-token rates cannot be compared without measuring your own audio. For gpt-4o-mini-tts users the audio output rate does rise, from $12 to $20 per million tokens, on the same text input rate. One inconsistency to watch: the text-to-speech guide we read on October 7 still presents gpt-4o-mini-tts as the current model with no deprecation notice. Text to speech guide The deprecations page is the one that carries dates, so plan from it.
Everything else on the calendar through February
The same page lists shutdowns that are not models or not October. None is new, but a checklist that stops at October 23 misses a working Evals or Agent Builder dependency.
| Date | What stops | Substitute |
|---|---|---|
| Oct 31, 2026 | Existing evals become read-only | None listed; the Evals section links a cookbook on moving to Promptfoo |
| Nov 30, 2026 | Evals dashboard and API, Agent Builder, v1/prompts and reusable prompt objects | Agents SDK or ChatGPT Workspace Agents; move prompt text into application code |
| Dec 1, 2026 | gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest | gpt-image-2.5-sunburst or -flare |
| Dec 11, 2026 | gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro snapshots; o3, o3-pro | GPT-5.6 Sol, Terra or Luna |
| Jan 20, 2027 | gpt-realtime, gpt-audio, gpt-4o-audio, gpt-4o-realtime and their mini variants | gpt-realtime-2.1, gpt-audio-1.5 and the 2.1-mini; the gpt-4o-mini-transcribe-2025-03-20 snapshot also moves to the -2025-12-15 snapshot |
| Feb 26, 2027 | whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize | gpt-live-transcribe or gpt-transcribe |
Source: OpenAI deprecations page.
The December 11 row catches people who moved to gpt-5 in August 2025 and think of it as current. It has 65 days left. The GPT-5.6 family those rows point at reached general availability in July; see our GPT-5.6 GA coverage for how the three tiers are positioned.
Copilot admins: October 19 is a different list under different rules
Keep the GitHub schedule in a separate column of your spreadsheet. GitHub is removing models from the Copilot product; it is not OpenAI retiring them from the API. gpt-5.5 and gpt-5.4 are still priced and absent from OpenAI’s deprecations page as of today. OpenAI pricing A team that calls gpt-5.5 directly has nothing to do on October 19.
On October 19, GitHub removes these models “across all GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions)”. GitHub changelog
| Removed from Copilot on Oct 19 | GitHub’s suggested alternative |
|---|---|
| GPT-5.5 | GPT-5.6 Sol |
| GPT-5.4 | GPT-5.6 Sol |
| GPT-5.4 mini | GPT-5.6 Luna |
| GPT-5 mini | GPT-5.6 Luna |
| Gemini 3.7 Flash | Gemini 3.8 Flash |
| Grok 4.5 | Grok 4.6 |
What happens to users on that day depends on a policy setting that dates from August. GitHub’s “Default availability for released models” policy for Copilot Business and Enterprise took effect on August 26: models an admin never explicitly configured follow the global default, and the global default is enabled. Default availability docs July 29 changelog The September post applies that to the retirements: the suggested alternatives “are automatically enabled for Copilot Enterprise and Copilot Business customers unless an administrator has turned off the global default or explicitly disabled the model”. If you turned the default off, you enable each replacement through its model policy in Copilot settings, or your developers lose GPT-5.5 on the 19th and gain nothing. GitHub changelog
Explicit choices survive either way. GitHub’s July post says a model you deliberately enabled or disabled “is never touched”, so an organization that blocked GPT-5.6 Sol in July for a review reason will still have it blocked when GPT-5.5 disappears. July 29 changelog Check the Models page under your enterprise’s AI controls or your organization’s Copilot settings: anything labelled as delegating to the default policy will follow the global switch, anything with an explicit setting will not. GitHub also says no action is needed to remove the retired models themselves. GitHub changelog
Two things the post does not cover and we could not confirm: whether any Copilot API or extension that passes a model name will return an error or silently reroute after the 19th, and which model the chat selector will fall back to for a user whose chosen model is gone. Test both with a pilot user before the date if either matters to you.
The dated version
| When | What to do |
|---|---|
| This week | Run the grep, search config and automation tools, and list every workload per retiring ID. Ask the account owner for OpenAI’s deprecation emails. |
| By Oct 12 | Copilot Business and Enterprise admins check the default-availability policy and the explicit setting on GPT-5.6 Sol, Luna, Gemini 3.8 Flash and Grok 4.6. |
| By Oct 16 | For each OpenAI workload, pick a replacement by test, not by the deprecations table alone. Set reasoning.effort explicitly. Record the cost on a day of real traffic. |
| Oct 19 | Copilot removals. |
| Oct 23 | OpenAI shutdowns, including every fine-tune on a retired base. Watch the scheduled jobs that run overnight. |
| Nov 21 | Earliest date GPT-5.6 Sol’s promotional rates can end. Re-check the bill. |
| Dec 11, Jan 6, Jan 20, Feb 26, Apr 1 | The rest of the list above. Put them in the same calendar now, while the grep results are fresh. |