GPT-4 shutdown on 23 October: the migration map

The GPT-4 shutdown lands on 23 October 2026, 32 days from today. On that date OpenAI removes gpt-4, gpt-3.5-turbo, gpt-4-turbo, o1, o1-pro, o3-mini, o4-mini, gpt-4.1-nano, gpt-4o-2024-05-13, gpt-image-1 and five families of fine-tuned models from the API. Below is OpenAI's full substitute table, what each move does to your per-token price, and the behaviour changes that break quietly.

• Shutdown date: 23 October 2026, announced on 22 April 2026.
• What happens: at shutdown a model is no longer accessible, so requests naming it stop working.
• Who is affected: any code, config or fine-tuned model that sends one of the IDs below.
• Sooner still: gpt-3.5-turbo-instruct, gpt-3.5-turbo-1106, babbage-002 and davinci-002 shut down on 28 September 2026, with gpt-5.6-terra as the substitute.

The GPT-4 shutdown migration map, model by model

This is the table from OpenAI's deprecations page, reproduced as published. Where a row shows a pipe, the first ID is the snapshot and the IDs after it are aliases that go with it.

Shutdown date	Model snapshot	Substitute model
October 23, 2026	gpt-3.5-turbo-0125 | gpt-3.5-turbo, gpt-3.5-turbo-completions	gpt-5.6-terra
October 23, 2026	gpt-4-0613 | gpt-4, gpt-4-0613-completions, gpt-4-completions	gpt-5.6-sol
October 23, 2026	gpt-4-1106-preview	gpt-5.6-sol
October 23, 2026	gpt-4-turbo | gpt-4-turbo-2024-04-09, gpt-4-turbo-completions	gpt-5.6-sol
October 23, 2026	gpt-4.1-nano | gpt-4.1-nano-2025-04-14	gpt-5.6-luna
October 23, 2026	gpt-4o-2024-05-13	gpt-5.6-sol
October 23, 2026	gpt-image-1	gpt-image-2
October 23, 2026	o1-2024-12-17 | o1	gpt-5.6-sol
October 23, 2026	o1-pro-2025-03-19 | o1-pro	gpt-5.6-sol (reasoning.mode: pro)
October 23, 2026	o3-mini-2025-01-31 | o3-mini	gpt-5.6-sol
October 23, 2026	ft-o4-mini-2025-04-16	gpt-5.6-terra
October 23, 2026	o4-mini-2025-04-16 | o4-mini	gpt-5.6-terra

OpenAI is also removing fine-tuned versions on the same date:

Shutdown date	Model snapshot	Recommended replacement base model
October 23, 2026	ft-gpt-3.5-turbo	gpt-5.6-terra
October 23, 2026	ft-gpt-4	gpt-5.6-sol
October 23, 2026	ft-gpt-4.1-nano-2025-04-14	gpt-5.6-luna
October 23, 2026	ft-babbage-002	gpt-5.6-terra
October 23, 2026	ft-davinci-002	gpt-5.6-terra

Two readings of this table matter. First, the alias does not save you: gpt-4 and the gpt-4-0613 snapshot go together. Second, nothing here is a like-for-like swap. Every text substitute is a GPT-5.6 reasoning model, and several of the models they replace, including gpt-4 and gpt-3.5-turbo, were not.

What each move does to your price

Here are the Standard per-million-token prices from OpenAI's pricing page for every row where both the old and new model are listed. Output tokens on the new models include reasoning tokens, which is the subject of the next section.

Old model	Old input / output	Substitute	New input / output	Change
gpt-4-0613	$30.00 / $60.00	gpt-5.6-sol	$4.00 / $20.00	Input 87% lower, output 67% lower
gpt-4-turbo-2024-04-09	$10.00 / $30.00	gpt-5.6-sol	$4.00 / $20.00	Input 60% lower, output 33% lower
gpt-4o-2024-05-13	$5.00 / $15.00	gpt-5.6-sol	$4.00 / $20.00	Input 20% lower, output 33% higher
o1	$15.00 / $60.00	gpt-5.6-sol	$4.00 / $20.00	Input 73% lower, output 67% lower
o1-pro	$150.00 / $600.00	gpt-5.6-sol (pro mode)	$4.00 / $20.00 per token, more tokens	Lower rate; pro mode uses more tokens
o3-mini	$1.10 / $4.40	gpt-5.6-sol	$4.00 / $20.00	Input 3.6x, output 4.5x
o4-mini	$1.10 / $4.40	gpt-5.6-terra	$2.00 / $12.00	Input 1.8x, output 2.7x
gpt-3.5-turbo-0125	$0.50 / $1.50	gpt-5.6-terra	$2.00 / $12.00	Input 4x, output 8x
gpt-4.1-nano	$0.10 / $0.40	gpt-5.6-luna	$0.20 / $1.20	Input 2x, output 3x
gpt-image-1 (image tokens)	$10.00 / $40.00	gpt-image-2	$8.00 / $30.00	Input 20% lower, output 25% lower

The arithmetic is simple division: gpt-4-0613 input falls from $30 to $4, which is 4 / 30 = 0.13, an 87% cut; gpt-3.5-turbo output rises from $1.50 to $12, which is 12 / 1.5 = 8x. gpt-4-1106-preview and the fine-tuned models have no Standard price on the page, so we have left them out rather than guess.

Two cautions on the Sol rows. GPT-5.6 Sol's current rates are promotional pricing available at least through 21 November 2026, so budget for them to change. And prompts over 272K input tokens are billed at 2x input and 1.5x output for the full request.

The honest summary: GPT-4, GPT-4 Turbo and o1 users get cheaper tokens. Teams on the cheap models, gpt-3.5-turbo, o3-mini, o4-mini and gpt-4.1-nano, pay more per token for a newer model, and the high-volume workloads usually sit exactly there. If the substitute is more model than a task needs, check whether gpt-5.6-luna, at $0.20 input and $1.20 output, passes your tests before accepting OpenAI's default. We looked at where token spend concentrates in LLM cost optimisation.

Migration gotchas that break quietly

Reasoning is on by default, and it bills as output

All three GPT-5.6 models default to medium reasoning effort and support none through max. Reasoning tokens are invisible in the response but billed as output tokens and occupy context. A gpt-3.5-turbo classifier moved to Terra with default settings will think before every answer, which adds latency and output cost the price table above does not show. For latency-critical tasks such as classification, OpenAI's guidance points to none; set the effort explicitly and measure.

The legacy Completions endpoint has no substitute

The table lists gpt-4-completions and gpt-3.5-turbo-completions, but the GPT-5.6 Sol model page marks v1/completions as not supported, and Terra and Luna are the same. Code calling the legacy Completions endpoint needs a new endpoint as well as a new model ID. If you are rewriting that call anyway, move it to the Responses API, which OpenAI's migration guide covers.

Fine-tuned models cannot be re-fine-tuned on the substitute

OpenAI recommends GPT-5.6 base models for the fine-tuned rows, yet the GPT-5.6 model pages list fine-tuning as not supported, and OpenAI is winding down the fine-tuning platform. Whatever your ft-gpt-4 or ft-gpt-3.5-turbo model learned has to be recovered with instructions and examples in the prompt, then proved against the old model's outputs. Start this one first, because it takes longest.

Jumping to GPT-6 Astra instead changes the API contract

Some teams will skip Sol and go straight to GPT-6 Astra. OpenAI's Astra migration notes list what breaks: no none reasoning effort, temperature, top_p and top_logprobs must be removed, and while Astra supports Chat Completions, tool calling requires the Responses API. Astra is also priced at $10 input and $50 output per million tokens.

Prompt caching now has a write charge

On GPT-5.6 and later, cache writes cost 1.25x the uncached input rate and reads cost 0.1x, and the cache lifetime setting moves from prompt_cache_retention to prompt_cache_options.ttl, whose only value is 30m. Older models had no write charge. Put stable instructions first so the write is reused, and keep the prefix above the 1,024-token minimum.

A two-week plan for an app pinned to a 2023-era model

1. Days 1-2, inventory. Search code, environment variables, config stores and feature flags for every model ID in the tables above, including ft: fine-tuned IDs and the -completions variants. Record the endpoint each call uses.
2. Days 3-4, build the test set. Pull a few hundred real inputs per call site from logs, with the current model's outputs as the baseline. If you grade with an LLM, correct for the known LLM-as-a-judge biases.
3. Days 5-7, swap and set effort. Change the model ID, set reasoning.effort explicitly, remove unsupported parameters, and move Completions calls to Responses.
4. Days 8-10, compare. Run the test set on the substitute and on the next tier down. Compare quality, latency and cost per call, including reasoning tokens.
5. Days 11-14, roll out. Ship behind a flag, watch error rates and spend, and remove the old ID from config before 23 October.

If the migration cannot finish in time, OpenAI's deprecations page notes that some developers may be able to provision dedicated capacity for continued access after a shutdown date, through its sales team. That is a bridge, not a plan. The same pattern is running across vendors this month; GitHub's Copilot model deprecation lands on 19 October. If you want a second pair of hands on the inventory or the test set, that is the work our AI engineering services team does.