OpenAI Shuts Down GPT-4, GPT-4 Turbo, GPT-3.5 Turbo, o1, o1-pro and o3-mini on October 23, 2026: the Replacements and What They Cost
On October 23, 2026, OpenAI turns off six older models in its API: GPT-4, GPT-4 Turbo, GPT-3.5 Turbo, o1, o1-pro and o3-mini. The date, and the model OpenAI recommends in place of each one, come from a single table on OpenAI’s deprecations page, under the announcement “2026-04-22: Legacy GPT model snapshots”, which opens: “To improve reliability and make it easier for developers to choose the right models, we are deprecating a set of older OpenAI models. Access to these models will be shut down on the dates below.” This post reads that table, and the model page of every model involved, as they stood on October 5, 2026, and puts a price on each move. It compares prices only: it says nothing about which model answers better, because nothing on these pages measures that and this site does not either.
The same shutdown is on this site’s deprecations calendar, which lists every announced date with the provider’s own words.
The six rows, as OpenAI prints them
The deprecations table has three columns: shutdown date, model snapshot (with its aliases) and substitute model. These are the six rows for the models this site tracks, copied as they read on October 5, 2026. The date and the substitute are identical to what this site recorded when it first read the table, on October 2.
- GPT-4:
October 23, 2026 | gpt-4-0613 | gpt-4 , gpt-4-0613-completions , gpt-4-completions | gpt-5.6-sol | - GPT-4 Turbo:
October 23, 2026 | gpt-4-turbo | gpt-4-turbo-2024-04-09 , gpt-4-turbo-completions | gpt-5.6-sol | - GPT-3.5 Turbo:
October 23, 2026 | gpt-3.5-turbo-0125 | gpt-3.5-turbo , gpt-3.5-turbo-completions | gpt-5.6-terra | - o1:
October 23, 2026 | o1-2024-12-17 | o1 | gpt-5.6-sol | - o1-pro:
October 23, 2026 | o1-pro-2025-03-19 | o1-pro | gpt-5.6-sol ( reasoning.mode: pro ) | - o3-mini:
October 23, 2026 | o3-mini-2025-01-31 | o3-mini | gpt-5.6-sol |
Five of the six point to the same model,
GPT-5.6 Sol. Only GPT-3.5 Turbo goes
elsewhere, to GPT-5.6 Terra. Note what
the aliases column implies: the plain names gpt-4, gpt-4-turbo,
gpt-3.5-turbo, o1, o1-pro and o3-mini stop working on the same
day as the dated snapshots, so code that never pinned a snapshot is
affected just the same.
The same table shuts down, on the same date, several models this site
does not track because it has no row for them: gpt-4-1106-preview and
gpt-4o-2024-05-13 (substitute gpt-5.6-sol), gpt-4.1-nano
(substitute gpt-5.6-luna), o4-mini (substitute gpt-5.6-terra) and
the image model gpt-image-1. Their prices are not covered here.
What the old models cost, and what the replacements cost
OpenAI’s general pricing page no longer lists any of the six models. Their own model pages still do. Each one prints a block headed “Text tokens”, “Per 1M tokens”, with a switch labelled “Batch API price”; on October 5, 2026 that switch was off, so the figures below are the standard prices, not the Batch ones. They were read by hand from each model page on October 5, 2026, and this site’s dataset carries them from that day, each with the line of its page quoted; the pricing table shows them among the deprecated models, and the deprecations calendar has the shutdown date.
| Model | Input / 1M | Output / 1M | Context window | Substitute |
|---|---|---|---|---|
| GPT-4 | $30.00 | $60.00 | 8,192 | GPT-5.6 Sol |
| GPT-4 Turbo | $10.00 | $30.00 | 128,000 | GPT-5.6 Sol |
| GPT-3.5 Turbo | $0.50 | $1.50 | 16,385 | GPT-5.6 Terra |
| o1 | $15.00 (cached $7.50) | $60.00 | 200,000 | GPT-5.6 Sol |
| o1-pro | $150.00 | $600.00 | 200,000 | GPT-5.6 Sol |
| o3-mini | $1.10 (cached $0.55) | $4.40 | 200,000 | GPT-5.6 Sol |
The two substitutes, from their own model pages on the same day:
- GPT-5.6 Sol: “Input $4.00 Cached input $0.40 Output $20.00”, with a “1,050,000 context window”. The same page says this is a promotion: “GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026.” OpenAI does not publish the price that follows, so this site does not show one and does not estimate one. Anyone budgeting past November 21 is budgeting against a number OpenAI has not given.
- GPT-5.6 Terra: “Input $2.00 Cached input $0.20 Output $12.00”, also with a “1,050,000 context window”. No promotion is mentioned.
Both pages add the same two rules: “Prompts with >272K input tokens are priced at 2x input and 1.5x output for the full request” and “Cache writes are billed at 1.25x the uncached input token rate.” Neither rule touches the comparison below, which uses short prompts and no cache. This site’s weekly check verified both prices again on the morning of October 5, 2026.
What the bill does, at one fixed workload
To put the six moves side by side, take one workload and keep it fixed: 10M input tokens and 2M output tokens a month, no cache hits, and prompts short enough that no long-context tier applies (all six old models have windows below 272,000 tokens anyway). The arithmetic is the one the cost calculator uses: tokens divided by a million, times the price per million, input and output added.
| Move | Old model | Substitute | Change |
|---|---|---|---|
| GPT-4 → GPT-5.6 Sol | $420.00 | $80.00 | $340.00 less, 5.25× cheaper |
| GPT-4 Turbo → GPT-5.6 Sol | $160.00 | $80.00 | $80.00 less, 2× cheaper |
| o1 → GPT-5.6 Sol | $270.00 | $80.00 | $190.00 less, 3.375× cheaper |
| o1-pro → GPT-5.6 Sol | $2,700.00 | $80.00 | $2,620.00 less, 33.75× cheaper |
| o3-mini → GPT-5.6 Sol | $19.80 | $80.00 | $60.20 more, about 4.04× dearer |
| GPT-3.5 Turbo → GPT-5.6 Terra | $8.00 | $44.00 | $36.00 more, 5.5× dearer |
Four moves lower the bill and two raise it. The two that raise it are the two cheapest old models: o3-mini, whose input price of $1.10 is under a third of Sol’s $4.00, and GPT-3.5 Turbo, whose $0.50 input is a quarter of Terra’s $2.00. For those two, the substitute OpenAI names is not the cheapest of its own models with the same context window, as the next section shows.
Two limits on these figures. First, they count tokens, not tasks: the same job can take a different number of tokens on another model, and reasoning models in particular spend tokens before they answer. The table says what the same token count costs; it does not predict the token count. Second, the Sol figure is a promotional price with no published end price, so the $80.00 holds at least through November 21, 2026 and is unknown after that.
Cheaper options with the substitute’s context window
If the reason to pick a model is its context window, the comparison that matters is with every model that offers at least as much. Both substitutes have 1,050,000 tokens. In this site’s dataset, seven current models with a verified price reach that, and all seven are OpenAI’s. At the same workload, cheapest first:
| Model | Input / 1M | Output / 1M | Monthly cost | Verified |
|---|---|---|---|---|
| GPT-6 Luna | $0.10 | $0.50 | $2.00 | October 5, 2026 |
| GPT-5.6 Luna | $0.20 | $1.20 | $4.40 | October 5, 2026 |
| GPT-6.1 Sol | $2.00 | $10.00 | $40.00 | October 5, 2026 |
| GPT-6 Sol | $2.00 | $10.00 | $40.00 | October 5, 2026 |
| GPT-5.6 Terra | $2.00 | $12.00 | $44.00 | October 5, 2026 |
| GPT-5.6 Sol | $4.00 | $20.00 | $80.00 | October 5, 2026 |
| GPT-6 Astra | $10.00 | $50.00 | $200.00 | October 5, 2026 |
All seven charge more above 272,000 input tokens per prompt; the figures above are the first band. GPT-5.6 Sol, the substitute named for five of the six shutdowns, is the second most expensive of the seven at this workload. GPT-6 Sol and GPT-6.1 Sol cost half of it on input and output. This is a price comparison and nothing else: OpenAI’s table names Sol and Terra, and whether another model does the same job is a question these pages do not answer.
One step down, at a context window of at least 1,000,000 tokens, other providers come in. The five cheapest at the same workload:
| Model | Context | Monthly cost | Condition | Verified |
|---|---|---|---|---|
| Qwen3.7-Flash | 1,000,000 | $0.56 | Singapore price; $1.80 with prompts over 32,000 tokens, $3.60 over 256,000 | September 17, 2026 |
| Gemini 2.5 Flash-Lite | 1,048,576 | $1.80 | none | October 5, 2026 |
| Qwen3.8-Flash | 1,000,000 | $2.44 | Singapore price | September 11, 2026 |
| Qwen3.8-Omni-Flash | 1,000,000 | $2.44 | Singapore price | September 28, 2026 |
| DeepSeek V4.1 Flash | 1,000,000 | $5.40 | peak price; $2.70 off-peak | October 5, 2026 |
Google gives Gemini 2.5 Flash-Lite’s limit as an input token limit of 1,048,576, which falls 1,424 short of the substitutes’ 1,050,000 and is why it is not in the first table. Alibaba prices its models by region and this site publishes the Singapore (“Scope: International”) price. DeepSeek charges half outside its peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday), and this site publishes the peak price as the list price. The pricing table has every other model.
What changes in the code, where OpenAI says so
The documentation states a few things about the move itself, and this site repeats only those.
- o1-pro needs a parameter. The substitute column reads “gpt-5.6-sol ( reasoning.mode: pro )”. The model page of GPT-5.6 Sol prints one price and says nothing about whether that mode is billed differently, so the comparison above uses the printed price.
- o1-pro already lived in one API. Its model page says “o1-pro is available in the Responses API only”. Code that calls it already uses the Responses API.
- The completions aliases go too. The aliases column lists
gpt-4-completions,gpt-4-0613-completions,gpt-4-turbo-completionsandgpt-3.5-turbo-completions, all shut down on the same date. - Fine-tuned models. The same announcement adds: “We are also
removing fine-tuned versions as below”, with
ft-gpt-4pointing togpt-5.6-solandft-gpt-3.5-turbotogpt-5.6-terraas the “recommended replacement base model”, both on October 23, 2026. Separately, the page says that “Inference on fine-tuned models will continue to be available until the base models are deprecated”. The model pages of GPT-4 and GPT-3.5 Turbo list fine-tuning as “Supported”; those of GPT-5.6 Sol and GPT-5.6 Terra list it as “Not supported”. - Older advice on the old pages. The model page of GPT-4 Turbo
still says “Today, we recommend using a newer model like GPT-4o”, and
the page of GPT-3.5 Turbo still says “use gpt-4o-mini in place of
GPT-3.5 Turbo”.
gpt-4o-2024-05-13is on the October 23 list.
This post does not go beyond what those pages state about the move; anything else about request format is for OpenAI’s documentation to say.
How this was read
Every figure in this post comes from OpenAI’s deprecations page and the model pages of the eight models involved, downloaded on October 5, 2026, or from this site’s dataset, where every row links to the page it was read on and the date. Prices are standard, per million tokens, in US dollars. The shutdown entry, with the provider’s own sentences, stays on the deprecations calendar until the date passes, and then moves to its list of models already shut down.
Frequently asked questions
When does OpenAI shut down GPT-4 and o1?
On October 23, 2026, according to OpenAI's deprecations page as read on October 5, 2026. The same date applies to GPT-4 Turbo, GPT-3.5 Turbo, o1-pro and o3-mini.
What replaces o1, o1-pro and o3-mini?
OpenAI names gpt-5.6-sol for all three, and for o1-pro it adds reasoning.mode: pro. GPT-5.6 Sol costs $4.00 per million input tokens and $20.00 per million output tokens, a promotional price that OpenAI says runs at least through November 21, 2026.
Does moving from GPT-4 to GPT-5.6 Sol cost more?
At the prices OpenAI printed on October 5, 2026, 10M input and 2M output tokens a month cost $420.00 on GPT-4 and $80.00 on GPT-5.6 Sol. OpenAI publishes no price for Sol after its promotion, so that figure can change.
What replaces GPT-3.5 Turbo, and what does it cost?
OpenAI names gpt-5.6-terra, at $2.00 per million input tokens and $12.00 per million output tokens. The same monthly workload goes from $8.00 on GPT-3.5 Turbo to $44.00 on GPT-5.6 Terra.
Drafted with AI assistance from the cited source and reviewed by Álvaro Lucero Rufino, who is responsible for what stays published here. Automated publishing was discontinued in September 2026; posts that no longer met that standard were removed rather than left up. This is original explanatory writing, not a reproduction — read the source above for the full story.