ModelsGPT-4o Guide: Three Snapshots, Three Prices, Two Shutdown Dates
GPT-4o is one name over three dated snapshots with different prices, output limits and retirement dates. Which one your code calls, and what to do before each deadline.
Most people think of GPT-4o as one model. In the API it is three dated snapshots behind one alias, and they do not behave, cost or retire the same way. If your code says gpt-4o, you are calling one of them, and which one decides your bill and your migration date.
That matters now because OpenAI has put deadlines on all three. gpt-4o-2024-05-13 shuts down on 23 October 2026. The other two, gpt-4o-2024-08-06 and gpt-4o-2024-11-20, follow on 14 January 2027.
What omni meant
OpenAI released GPT-4o on 13 May 2024; the "o" stands for omni. The system card describes a single network trained end to end across text, vision and audio, rather than separate models stitched together. It reports audio responses in as little as 232 milliseconds, averaging 320, and says GPT-4o matched GPT-4 Turbo on English text and code while being much faster and 50% cheaper in the API.
Through the standard text API, most of that omni story is invisible. OpenAI's model page lists text and image input with text output. The audio and realtime capabilities shipped as separate models, such as the gpt-4o-audio and gpt-4o-realtime families, which have their own deprecation schedule ending in early 2027.
OpenAI lists the main model with a 128,000-token context window, up to 16,384 output tokens and an October 2023 knowledge cutoff.
Three snapshots, three behaviours
The dated snapshots are frozen versions. They exist so that an application tested against one does not change behaviour when OpenAI improves the model. Here is how the three differ:
gpt-4o-2024-05-13is the launch version. OpenAI's pricing page lists it at $5 per million input tokens and $15 per million output, double the later input rate, and OpenRouter lists its output limit at 4,096 tokens.gpt-4o-2024-08-06cut the price to $2.50 and $10 and raised output to 16,384 tokens. OpenRouter's listing ties it to improved structured outputs, the ability to supply a JSON schema inresponse_formatand have the response conform to it.gpt-4o-2024-11-20keeps the same price and limits. OpenRouter's listing describes it as tuned for more natural, engaging creative writing.
The practical upshot is that a team still pinned to the May snapshot has been paying double for input since August 2024 with a quarter of the output room. That is worth fixing regardless of any deadline.
If you rely on schema-constrained JSON, structured outputs and JSON mode explains the difference between the older JSON mode, which only promises valid JSON, and schema enforcement, which promises your shape.
The alias, and why to pin
OpenAI's model page lists gpt-4o-2024-08-06 as the default snapshot, so that is what the bare gpt-4o alias currently resolves to. Aliases can move, which is convenient for chat applications and risky for anything that parses output.
Pin a dated snapshot in production, and record which one in your prompt version history. When you do change snapshots, run your regression set first. A move from August to November is small on paper, but it was explicitly a change to writing style, and style changes break parsers.
OpenRouter's catalogue mirrors the same split: openai/gpt-4o plus the three dated IDs, and a :batch endpoint that is the same model on an asynchronous queue at half price, $1.25 and $5, matching OpenAI's own Batch API rate.
The retirement timeline
GPT-4o has been leaving in stages:
- 13 February 2026. Removed from ChatGPT. The API was unaffected.
- 17 February 2026. The
chatgpt-4o-latestAPI snapshot shut down. - 23 October 2026.
gpt-4o-2024-05-13shuts down, withgpt-5.6-solas the named replacement. - 14 January 2027.
gpt-4o-2024-08-06andgpt-4o-2024-11-20shut down, withgpt-6-lunanamed as the replacement. Fine-tuned GPT-4o models follow the same date.
OpenAI's deprecation table names the dated snapshots rather than the bare alias. Since the alias points at the August snapshot, plan as though gpt-4o stops on 14 January 2027 unless OpenAI says otherwise.
GPT-4o or GPT-4o mini
The sibling decision still comes up for teams buying time before migrating. GPT-4o mini lists at $0.15 and $0.60, less than a sixteenth of GPT-4o's price, with the same 128,000-token window and the same 14 January 2027 shutdown.
Use full GPT-4o where answers need judgement: nuanced writing, multi-step instructions, ambiguous extraction. Use mini for classification, routing and short structured extraction, where its errors are cheap to catch. If the two disagree often on your own data, that disagreement is the signal for which tasks need the larger model.
Choosing a replacement
OpenAI names gpt-6-luna for the later snapshots, but the right replacement depends on what GPT-4o was doing for you. A chat assistant, a JSON extractor and a vision pipeline will each favour a different successor.
Build a small fixed set of real prompts per use case, run the candidates against it, and compare cost per completed task rather than price per token. How to benchmark LLMs on your own work has the setup, and model deprecation and migration covers sequencing a move when several snapshots retire on different dates.
Common questions
Which snapshot does the gpt-4o alias use?
OpenAI's model page lists gpt-4o-2024-08-06 as the default snapshot, so that is what the bare alias resolves to at the time of writing. Pin a dated snapshot in production so behaviour does not change underneath you.
When is GPT-4o being retired from the API?
In two steps. gpt-4o-2024-05-13 shuts down on 23 October 2026, and gpt-4o-2024-08-06 and gpt-4o-2024-11-20 shut down on 14 January 2027, according to OpenAI's deprecations page.
Why is gpt-4o-2024-05-13 more expensive than the newer snapshots?
It kept its launch price of $5 input and $15 output per million tokens. The August 2024 snapshot cut that to $2.50 and $10 and raised the output limit, so staying on the May version costs more for less.
