OpenAI GPT Model Release Timeline - Model Lineage, ChatGPT and Codex Milestones, and Platform Availability

First Published: 2026-06-19
Last Updated: 2026-07-13

This article is a companion to my Anthropic Claude Model Release Timeline, extending the same history-and-timeline format to OpenAI's GPT models. Where the Claude timeline tracks a single, cleanly tiered model family, OpenAI's story is broader: a line of GPT base models that branched into multimodal "omni" models and a separate family of o-series reasoning models, before converging into the unified GPT-5 generation - and, along the way, becoming the foundation for a wide range of products and services such as ChatGPT, Codex, DALL·E, Sora, and Whisper.

In this article I build a complete release timeline for OpenAI's GPT model lineage, from GPT-1 in 2018 through the GPT-5 generation in 2025 and 2026, with the major products and services built on top of those models woven into the same chronology. I summarize the GPT model family tree, the full chronology of major releases with links to the official openai.com announcements, the evolution of capabilities (long context, function calling, JSON mode, vision, voice and the Realtime API, reasoning, native computer use, and agents), the products and services built on GPT, the availability of GPT models across the OpenAI API, Azure OpenAI in Microsoft Foundry, Amazon Bedrock, Microsoft Copilot, and GitHub Copilot, and the lifecycle of deprecated and retired models as listed on OpenAI's official model deprecations page.

Companion articles on hidekazu-konishi.com:

Overview - From OpenAI's Founding to GPT-5.6

OpenAI was founded as a non-profit artificial-intelligence research laboratory on December 11, 2015, with the stated mission of ensuring that artificial general intelligence benefits all of humanity.
The first Generative Pre-trained Transformer (GPT) arrived in June 2018 as a research model, followed by GPT-2 in 2019 and GPT-3 in 2020. GPT-3 was the first model OpenAI commercialized broadly, through the OpenAI API private beta in June 2020.
The inflection point came on November 30, 2022, when OpenAI released ChatGPT as a free research preview built on the GPT-3.5 series. ChatGPT made conversational GPT models accessible to a general audience and triggered the wave of generative-AI adoption that followed.

From there, the GPT line advanced quickly. GPT-4 (March 14, 2023) added image input and far stronger reasoning; GPT-4 Turbo (announced at the first DevDay in November 2023) introduced a 128K-token context window, JSON mode, and the Assistants API; and GPT-4o (May 2024) folded text, audio, and image into a single "omni" model, joined by the smaller GPT-4o mini.
In parallel, OpenAI opened a second model family: the o-series of reasoning models that "think" before answering, beginning with o1-preview in September 2024 and continuing through o3 and o4-mini in April 2025.

On August 7, 2025, the two lines converged into GPT-5, a unified system that pairs a fast model with a deeper reasoning model ("GPT-5 thinking") behind a real-time router that decides how much to think on each request. The GPT-5 generation has advanced rapidly since: GPT-5.1 (November 2025), GPT-5.2 (December 2025), GPT-5.4 (March 2026), GPT-5.5 (April 23, 2026, with GPT-5.5 Instant becoming the default model in ChatGPT on May 5, 2026), and GPT-5.6 (released publicly on July 9, 2026, roughly two weeks after a limited preview of GPT-5.6 Sol began on June 26, 2026), which is the current frontier generation as of the time of writing. GPT-5.6 also introduced a new naming scheme - Sol, Terra, and Luna - in which the number identifies the model generation and the named tiers identify durable capability tiers that can advance on their own cadence.
Throughout this evolution, GPT models have also been the foundation for a growing family of products - ChatGPT, the Codex coding agent, the DALL·E image models, the Sora video model, Whisper speech recognition, and the Assistants, Responses, and Realtime APIs - which this timeline tracks alongside the models themselves.

The scope of this article is limited to publicly announced models, products, and capabilities documented on openai.com and the official OpenAI and Microsoft documentation. I do not include unreleased or rumored future models. Pricing changes frequently and is deliberately omitted in favor of links to OpenAI's official pricing page, and I do not reproduce benchmark scores.

This timeline primarily references the following sources.

GPT Model Family Tree (Base / Multimodal / Reasoning / Unified GPT-5 / Codex)

Unlike Anthropic's cleanly tiered Opus / Sonnet / Haiku family, OpenAI's models are best understood as several lineages that branched apart and then converged. The diagram below organizes them into five lanes: the base GPT models, the multimodal "omni" models, the o-series reasoning models, the unified GPT-5 generation, and the Codex coding models.

The base GPT lane is the original line of general-purpose text models: GPT-1 (2018), GPT-2 (2019), GPT-3 (2020), the GPT-3.5 series that powered the first ChatGPT (2022), GPT-4 (2023), GPT-4 Turbo (2023), and the API-only GPT-4.1 family (2025). GPT-4.5 (2025) was a short-lived research-preview member of this lane.

The multimodal "omni" lane begins with GPT-4o (2024), the first OpenAI model designed from the start to handle text, audio, and image in one model, joined by the cost-efficient GPT-4o mini. These models brought native multimodality to the GPT line before it merged into GPT-5.

The reasoning (o-series) lane is the family of models trained to produce an internal chain of thought before answering: o1-preview and o1-mini (2024), o1 (2024), o3-mini (2025), and o3 and o4-mini (2025). This lane introduced the "thinking" paradigm that GPT-5 later absorbed.

The unified GPT-5 lane is where the base, multimodal, and reasoning lines converge. GPT-5 (August 2025) is a single system with a fast model, a deeper "thinking" reasoning model, and a router that chooses between them; subsequent releases GPT-5.1, GPT-5.2, GPT-5.4, GPT-5.5, and GPT-5.6 continued this design. Through GPT-5.5, the generation used Instant, Thinking, and Pro variants (note that the Instant and Thinking variants were not always versioned in lock-step - for example, the March 2026 launch shipped GPT-5.3 Instant alongside GPT-5.4 Thinking and GPT-5.4 Pro). GPT-5.6 (July 2026) then replaced this scheme with named capability tiers - GPT-5.6 Sol (the frontier tier), GPT-5.6 Terra (a balance of intelligence and cost), and GPT-5.6 Luna (the fastest, most cost-efficient tier) - in which the number identifies the generation and the tier names are durable tiers that can advance on their own cadence.

The Codex coding lane is a special case. The original Codex (2021) was a GPT-3 descendant fine-tuned on code; it was retired in 2023. OpenAI later revived the Codex name for an entirely new coding agent (2025), backed by dedicated coding models such as codex-1 and the GPT-5-Codex line, which derive from the o-series and GPT-5 generations rather than the 2021 model. Since GPT-5.5 (April 2026), the mainline frontier models have shipped directly in Codex.

Timeline of Major Releases

Here is the chronological timeline of OpenAI GPT model releases and the major products, services, and platform milestones built on them. Each row links to the official openai.com announcement or the equivalent official source where one is available.

* The table can be sorted by clicking on the column names.

Date	Event
2015-12-11	OpenAI is founded as a non-profit AI research laboratory. Source: Introducing OpenAI.
2018-06-11	GPT-1 is published - the first Generative Pre-trained Transformer, introducing the unsupervised pre-training plus supervised fine-tuning recipe. Source: Improving language understanding with unsupervised learning.
2019-02-14	GPT-2 is announced with a staged release; the full 1.5B-parameter model followed in November 2019. Source: Better language models and their implications.
2020-06-11	GPT-3 is released through the OpenAI API (private beta) - the first GPT model OpenAI commercialized broadly. Source: OpenAI API.
2021-01-05	DALL·E and CLIP are introduced, bringing text-to-image generation to the GPT family of research. Source: DALL·E: Creating images from text.
2021-06-29	GitHub Copilot launches in technical preview, powered by the original OpenAI Codex model. Source: Introducing GitHub Copilot.
2021-08-10	OpenAI Codex is announced (private beta) - a GPT-3 descendant fine-tuned on code that translates natural language into code. This is the original Codex, later retired (see Deprecations).
2022-04-06	DALL·E 2 is announced as a research preview, with higher-resolution and more realistic image generation. Source: DALL·E 2.
2022-09-21	Whisper is released as an open-source automatic speech recognition model. Source: Introducing Whisper.
2022-11-30	ChatGPT is launched as a free research preview, built on the GPT-3.5 series. Source: Introducing ChatGPT.
2023-02-01	ChatGPT Plus is introduced, the first paid ChatGPT subscription tier. Source: Introducing ChatGPT Plus.
2023-03-01	The Chat Completions API (`gpt-3.5-turbo`) and the Whisper API are released for developers - the same model that powers the ChatGPT product, now available in the API. Source: Introducing ChatGPT and Whisper APIs.
2023-03-14	GPT-4 is announced - a large multimodal model accepting text and image input (image input shipped first as a limited preview and broadened over 2023-2024), with 8K- and 32K-token context variants. Source: GPT-4.
2023-06-13	Function calling is added to the Chat Completions API, letting models output structured arguments to call developer-defined functions. Source: Function calling and other API updates.
2023-08-28	ChatGPT Enterprise is announced, adding enterprise security, longer context, and administrative controls. Source: Introducing ChatGPT Enterprise.
2023-09-25	ChatGPT gains voice and image input ("see, hear, and speak"). Source: ChatGPT can now see, hear, and speak.
2023-10-19	DALL·E 3 becomes available in ChatGPT (announced September 20, 2023), integrating image generation directly into the chat experience. Source: DALL·E 3.
2023-11-06	First DevDay: GPT-4 Turbo, JSON mode, the Assistants API, and GPTs are announced. GPT-4 Turbo introduced a 128K-token context window. Source: New models and developer products announced at DevDay.
2024-01-10	The GPT Store launches, alongside ChatGPT Team. Source: Introducing the GPT Store.
2024-05-13	GPT-4o ("omni") is announced - a single model handling text, audio, and image input and output, with a 128K-token context window. Source: Hello GPT-4o.
2024-07-18	GPT-4o mini is announced, a small, cost-efficient multimodal model that replaced GPT-3.5 Turbo in ChatGPT. Source: GPT-4o mini: advancing cost-efficient intelligence.
2024-09-12	o1-preview and o1-mini are announced - OpenAI's first reasoning models, trained to think before responding. Source: Introducing OpenAI o1-preview.
2024-10-01	The Realtime API enters public beta (DevDay 2024), enabling low-latency speech-to-speech apps. Source: Introducing the Realtime API.
2024-10-31	ChatGPT search is launched, adding live web results to ChatGPT. Source: Introducing ChatGPT search.
2024-12-05	The full o1 model and ChatGPT Pro are launched. o1 became available in the API on December 17, 2024. Source: Introducing ChatGPT Pro.
2024-12-09	Sora launches publicly (text-to-video), after its February 15, 2024 research-preview announcement. Source: Sora is here.
2025-01-23	Operator is introduced - a research-preview agent that uses a computer (browser) on the user's behalf. Source: Introducing Operator.
2025-01-31	o3-mini is released in ChatGPT and the API - the first reasoning model offered to free ChatGPT users. Source: OpenAI o3-mini.
2025-02-02	Deep research is launched in ChatGPT - an agent that performs multi-step web research. Source: Introducing deep research.
2025-02-27	GPT-4.5 is released as a research preview - OpenAI's largest non-reasoning chat model; later deprecated (see Deprecations). Source: Introducing GPT-4.5.
2025-03-11	The Responses API and the open-source Agents SDK are introduced, with built-in web search, file search, and computer-use tools. Source: New tools for building agents.
2025-03-25	GPT-4o native image generation is released in ChatGPT, replacing DALL·E 3 as the default image generator (later exposed in the API as `gpt-image-1`). Source: Introducing 4o image generation.
2025-04-14	GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano are released in the API with a 1-million-token context window. Source: Introducing GPT-4.1 in the API.
2025-04-16	o3 and o4-mini are announced - the first reasoning models able to agentically use every tool in ChatGPT. The Codex CLI open-source coding agent launched the same week. Source: Introducing OpenAI o3 and o4-mini.
2025-05-16	Codex (the cloud software-engineering agent) launches in ChatGPT as a research preview, powered by the `codex-1` model (a version of o3 optimized for software engineering). Source: Introducing Codex.
2025-07-17	ChatGPT agent is introduced, unifying Operator-style computer use and deep research into a single agent mode. Source: Introducing ChatGPT agent.
2025-08-07	GPT-5 is announced - a unified system with a fast model, a deeper "GPT-5 thinking" reasoning model, and a real-time router, in `gpt-5`, `gpt-5-mini`, and `gpt-5-nano` API sizes. It became the ChatGPT default. Source: Introducing GPT-5.
2025-08-28	The Realtime API reaches general availability with the `gpt-realtime` speech-to-speech model. Source: Introducing gpt-realtime and Realtime API updates.
2025-09-15	GPT-5-Codex is released, a version of GPT-5 optimized for agentic coding and used as the default model in Codex. Source: Introducing upgrades to Codex.
2025-09-30	Sora 2 is announced, OpenAI's flagship video-and-audio generation model, alongside a new Sora app. Source: Sora 2.
2025-10-06	DevDay 2025: Codex reaches general availability, and AgentKit and the Apps SDK are announced. Source: Codex is now generally available.
2025-11-12	GPT-5.1 is released (GPT-5.1 Instant and GPT-5.1 Thinking), adding adaptive reasoning and a warmer default tone. Source: GPT-5.1: A smarter, more conversational ChatGPT.
2025-12-11	GPT-5.2 is released, the next flagship in the GPT-5 family, aimed at professional knowledge work. Source: Introducing GPT-5.2.
2026-03-05	GPT-5.4 is released (with GPT-5.3 Instant, GPT-5.4 Thinking, and GPT-5.4 Pro) - the first mainline model with native computer-use, a 1-million-token context option in the API and Codex, and tool search. Model snapshot `gpt-5.4-2026-03-05`. Source: Introducing GPT-5.4.
2026-03-17	GPT-5.4 mini and GPT-5.4 nano are released, bringing GPT-5.4-class capability to smaller, high-volume models. Source: Introducing GPT-5.4 mini and nano.
2026-03-24	The discontinuation of the Sora app is announced. The Sora web and app experiences ended on April 26, 2026, and the Sora API and Sora 2 models are scheduled to shut down on September 24, 2026 (see Deprecations). Source: What to know about the Sora discontinuation.
2026-04-21	gpt-image-2 (ChatGPT Images 2.0) is released - the first OpenAI image model with thinking capabilities - replacing the DALL·E line as the default image generator ahead of the DALL·E 2 and DALL·E 3 retirement on May 12, 2026. Source: Introducing ChatGPT Images 2.0.
2026-04-23	GPT-5.5 is released in ChatGPT and Codex, with GPT-5.5 Pro for the most demanding work; it became available in the API on April 24, 2026. Source: Introducing GPT-5.5.
2026-05-05	GPT-5.5 Instant becomes the default model in ChatGPT for all users, with reduced hallucinations and improved personalization. Source: GPT-5.5 Instant.
2026-06-26	GPT-5.6 Sol is previewed to a small group of trusted partners under a new U.S. government pre-release review process, ahead of its public release on July 9, 2026. Source: Previewing GPT-5.6 Sol.
2026-07-08	GPT-Live-1 and GPT-Live-1 mini are announced - a new generation of full-duplex voice models that listen and speak at the same time, replacing Advanced Voice Mode in ChatGPT, with API availability planned. Source: Introducing GPT-Live.
2026-07-09	GPT-5.6 is released publicly - the current frontier generation as of the time of writing - in three named capability tiers (GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna) across ChatGPT, Codex, and the API, adding programmatic tool calling and a multi-agent orchestration beta to the Responses API. The same day, OpenAI announced the ChatGPT Work agent and a new ChatGPT desktop app that merges Codex into ChatGPT. Source: GPT-5.6: Frontier intelligence that scales with your ambition.

There may be slight variations in the dates in this timeline due to differences between the original announcement, general availability in the OpenAI API, and availability in ChatGPT or on cloud platforms. Where these differ, the dates above prioritize the original OpenAI announcement.

The content posted here is limited to major releases and milestones considered essential for understanding the GPT model family and the products built on it.
In other words, please note that the items on this timeline are not all of OpenAI's updates, but representative releases that I have picked out.

Capability Evolution

Beyond the model release dates themselves, OpenAI has shipped a steady stream of capabilities that apply across multiple GPT models. The table below tracks when each major capability first became available.

* The table can be sorted by clicking on the column names.

Capability	First Introduced	First Model / Surface	Notes
OpenAI API	2020-06-11	GPT-3	Initial private-beta developer access.
Chat Completions API	2023-03-01	gpt-3.5-turbo	Chat-oriented API powering ChatGPT-style apps.
Vision (image input)	2023-03-14	GPT-4	Multimodal image understanding (broad availability rolled out through 2023-2024).
Function calling	2023-06-13	gpt-4-0613 / gpt-3.5-turbo-0613	Structured tool/function invocation from the API.
Voice and image in ChatGPT	2023-09-25	GPT-4	Spoken conversations and image input in the ChatGPT product.
128K context window	2023-11-06	GPT-4 Turbo	Large context for long documents.
JSON mode	2023-11-06	GPT-4 Turbo	Guaranteed valid-JSON output.
Assistants API	2023-11-06	DevDay 2023	Stateful assistant primitive with tools and retrieval.
Native "omni" multimodality	2024-05-13	GPT-4o	Single model across text, audio, and image.
Reasoning ("thinking")	2024-09-12	o1-preview	Internal chain of thought before answering.
Realtime API (speech-to-speech)	2024-10-01	gpt-4o-realtime	Low-latency voice; GA with gpt-realtime on 2025-08-28.
Responses API and Agents SDK	2025-03-11	Platform	Agent-oriented API primitive with built-in tools and orchestration.
1M-token context window	2025-04-14	GPT-4.1	Million-token context in the API.
Agentic tool use in reasoning	2025-04-16	o3 / o4-mini	Reasoning models using every ChatGPT tool agentically.
Unified system with router	2025-08-07	GPT-5	Automatic routing between fast and thinking models.
Native computer use (general model)	2026-03-05	GPT-5.4	First mainline model with built-in computer-use.
Tool search	2026-03-05	GPT-5.4	Efficient handling of large tool sets in the API.
Thinking in image generation	2026-04-21	gpt-image-2	First OpenAI image model that reasons and plans before generating.
Full-duplex voice	2026-07-08	GPT-Live-1 / GPT-Live-1 mini	Listens and speaks simultaneously; replaced Advanced Voice Mode in ChatGPT (API planned).
Programmatic tool calling	2026-07-09	GPT-5.6 (Responses API)	Model-written programs orchestrate tool calls in a hosted runtime.
Multi-agent orchestration (beta)	2026-07-09	GPT-5.6 (Responses API)	Coordinates concurrent subagents within a single request.

Products and Services Built on GPT

A defining feature of OpenAI compared with a pure model provider is the breadth of products built directly on its GPT models. The diagram below maps the major products and services to the model families that power them, and the subsections that follow describe each in turn.

ChatGPT

ChatGPT, launched on November 30, 2022, is the product that brought GPT models to a mass audience. It began as a free research preview on GPT-3.5, gained the paid ChatGPT Plus tier in February 2023, and added GPT-4, plugins and browsing, ChatGPT Enterprise (August 2023), voice and image input (September 2023), the GPT Store (January 2024), memory, ChatGPT search (October 2024), and the ChatGPT Pro tier (December 2024). Over time it grew from a chat box into an agent platform with Operator, deep research, and the unified ChatGPT agent (2025). In July 2026, OpenAI took this further with ChatGPT Work - an agent for research and finished deliverables with built-in Codex - and a new ChatGPT desktop app for macOS and Windows that combines Chat, Work, and Codex in a single app (the previous desktop app was renamed ChatGPT Classic). Each major model generation - GPT-4, GPT-4o, the o-series, and the GPT-5 generation through GPT-5.6 - has shipped in ChatGPT alongside the API.

Codex (the original model and the new agent)

The Codex name refers to two very different things. The original Codex (announced August 10, 2021) was a GPT-3 descendant fine-tuned on source code; it translated natural language into code and was the model behind the first generation of GitHub Copilot. OpenAI deprecated and shut it down in March 2023, advising developers that its newer general-purpose GPT models were better at coding.

OpenAI later revived the name for an entirely new, unrelated Codex coding agent (2025): first the open-source Codex CLI terminal agent (April 2025), then a cloud-based software-engineering agent in ChatGPT (May 2025) powered by the codex-1 model, which reached general availability at DevDay in October 2025. The new Codex initially ran on a rapid succession of dedicated coding models - codex-mini, GPT-5-Codex, and the GPT-5.1-Codex, GPT-5.2-Codex, and GPT-5.3-Codex line - that derive from the o-series and GPT-5 generations rather than the 2021 model; since GPT-5.5 (April 2026), the mainline frontier models have shipped directly in Codex. In July 2026, the Codex app was merged into the new ChatGPT desktop app, while Codex remains available as the CLI and IDE integrations. In short, "Codex" today means a coding agent surface within ChatGPT, the CLI, and IDEs, not the retired 2021 code model.

Image, Audio, and Video

OpenAI's generative-media models also sit on top of, or alongside, the GPT line. DALL·E (January 2021), DALL·E 2 (April 2022), and DALL·E 3 (in ChatGPT from October 2023) covered text-to-image generation, until GPT-4o native image generation (March 2025, later exposed in the API as gpt-image-1) folded image generation into the multimodal GPT line; gpt-image-2 (ChatGPT Images 2.0, April 21, 2026) succeeded it as the default image model, and the DALL·E 2 and DALL·E 3 API models were retired on May 12, 2026. Whisper (September 2022) provided open-source speech recognition. For voice, Advanced Voice Mode (built on GPT-4o) and later the gpt-realtime models brought natural spoken conversation, and the full-duplex GPT-Live-1 models (July 2026) replaced Advanced Voice Mode in ChatGPT. For video, Sora was announced as a research preview in February 2024, launched publicly in December 2024, and was succeeded by Sora 2 in September 2025; OpenAI then announced the discontinuation of the Sora app in March 2026, ending the web and app experiences on April 26, 2026, with the Sora API scheduled to shut down on September 24, 2026.

Developer Platform and Agents

For developers, the GPT models are exposed through a platform that has steadily moved from simple completions toward agents. Milestones include the Chat Completions API (March 2023), function calling (June 2023), JSON mode and the Assistants API (DevDay 2023), the Realtime API (October 2024), and the Responses API together with the open-source Agents SDK and built-in tools (March 2025), followed by AgentKit (DevDay 2025) and, with GPT-5.6 (July 2026), programmatic tool calling and a multi-agent orchestration beta in the Responses API. On the product side, Operator (January 2025), deep research (February 2025), the unified ChatGPT agent (July 2025), and ChatGPT Work (July 2026) turned GPT models into agents that can browse, research, and operate software on a user's behalf.

Availability Across Platforms

GPT models are available through OpenAI's own API and ChatGPT, and through Microsoft's platforms - most directly Azure OpenAI (now part of Microsoft Foundry), and embedded in products such as Microsoft Copilot and GitHub Copilot. Since June 2026, OpenAI models are also generally available on Amazon Bedrock (described after the table below), ending the period in which Azure was the only cloud platform offering them. The table below summarizes when major model generations arrived on the OpenAI API and on Azure OpenAI / Microsoft Foundry.

* The table can be sorted by clicking on the column names.

Model / Generation	OpenAI API	Azure OpenAI in Microsoft Foundry
GPT-3	2020-06-11	2021-11 (preview)
GPT-3.5 / ChatGPT	2023-03-01	2023-03-09 (preview)
GPT-4	2023-03-14	2023-03-21 (preview)
GPT-4o	2024-05-13	2024-05-13
GPT-4o mini	2024-07-18	2024-07-18
o1	2024-12-17	2024-12
o3-mini	2025-01-31	2025-01-31
GPT-4.1	2025-04-14	2025-04
o3 / o4-mini	2025-04-16	2025-04
GPT-5 family	2025-08-07	2025-08
GPT-5.5	2026-04-24	2026-04
GPT-5.6 (Sol / Terra / Luna)	2026-07-09	2026-07-09

For dates shown at month precision, the exact day and regional rollout can vary; the Azure OpenAI dates reflect Microsoft's "What's new" documentation, which groups some releases by month. For the most current availability and exact regional rollouts, refer to the Azure OpenAI in Microsoft Foundry "What's new" page and the OpenAI models documentation.

Microsoft first offered OpenAI models through the Azure OpenAI Service, announced in November 2021 and made generally available on January 16, 2023. Its companion development platform was subsequently rebranded twice: Azure AI Studio became Azure AI Foundry in November 2024, then Microsoft Foundry in November 2025, with the Azure OpenAI models now offered within it as "Azure OpenAI in Foundry Models." Beyond the Azure platform, GPT-4 powered the new Bing (later rebranded Microsoft Copilot) from February 7, 2023 (Microsoft confirmed it had been running on GPT-4 the following month), and GitHub Copilot - launched in technical preview on June 29, 2021 and made generally available on June 21, 2022 - was originally powered by the original OpenAI Codex model and later moved to newer GPT and Codex models.

In June 2026, OpenAI models arrived on AWS for the first time: GPT-5.5, GPT-5.4, and Codex became generally available on Amazon Bedrock, offered through the Responses API with a 272K-token context window and pricing matching OpenAI's first-party rates, with GPT-5.4 also available in AWS GovCloud (US-West). As of the time of writing, GPT-5.6 has not yet been announced on Amazon Bedrock. Source: GPT-5.5, GPT-5.4, and Codex from OpenAI are now generally available on Amazon Bedrock.

Deprecations and Model Lifecycle

OpenAI publishes deprecation and shutdown dates for its models on its official deprecations page. A model is first marked as deprecated, with a recommended replacement, and then has a shutdown (retirement) date after which API calls to that model identifier fail. The table below lists notable model retirements drawn from the official OpenAI deprecations page. Note that OpenAI publishes some shutdowns as "no earlier than" dates (the earliest scheduled retirement, which can slip), so treat the shutdown dates below as the earliest scheduled day rather than a guaranteed exact one.

The Recommended Replacement column reflects the successor currently recommended on OpenAI's deprecations page as of the time of writing; OpenAI updates these recommendations as newer models are released, so for older deprecations the listed replacement is generally not the one suggested at the time of the original announcement.

* The table can be sorted by clicking on the column names.

Announced	Shutdown	Deprecated Model(s)	Recommended Replacement
2023-03-20	2023-03-23	Original Codex: `code-davinci-002`, `code-davinci-001`, `code-cushman-002`, `code-cushman-001`	`gpt-4o`
2023-06-13	2024-06-13	`gpt-4-0314` (original GPT-4 snapshot)	`gpt-4o`
2023-07-06	2024-01-04	GPT-3 base and InstructGPT text models (`davinci`, `curie`, `babbage`, `ada`, `text-davinci-003`, etc.) and first-generation embeddings	`gpt-3.5-turbo-instruct` / `davinci-002` / `text-embedding-3-small`
2024-06-06	2024-12-06	`gpt-4-vision-preview`, `gpt-4-1106-vision-preview`	`gpt-4o`
2024-06-06	2025-06-06	`gpt-4-32k` and its snapshots	`gpt-4o`
2025-04-14	2025-07-14	`gpt-4.5-preview`	`gpt-4.1`
2025-11-14	2026-05-12	DALL·E image models: `dall-e-2`, `dall-e-3`	`gpt-image-2` / `gpt-image-1` / `gpt-image-1-mini`
2025-11-17	2026-02-12	`codex-mini-latest`	`gpt-5-codex-mini`
2025-11-18	2026-02-17	`chatgpt-4o-latest`	`gpt-5.1-chat-latest`
2026-03-24	2026-09-24	Sora 2 video models (`sora-2`, `sora-2-pro`, and their snapshots) and the Videos API	(none listed)
2026-04-22	2026-07-23	Preview and research models, including `computer-use-preview` and `o3-deep-research`	`gpt-5.5` / `gpt-5.4-mini`
2026-04-22	2026-10-23	Legacy chat models and snapshots, including `gpt-3.5-turbo`, GPT-4-era and GPT-4o snapshots, `o1`, `o1-pro`, `o3-mini`, and `o4-mini`	`gpt-5.5` / `gpt-5.4-mini` / `gpt-5.5-pro`
2026-06-11	2026-12-11	Original GPT-5 snapshots (`gpt-5`, `gpt-5-mini`, `gpt-5-nano`, `gpt-5-pro`) and o3 snapshots (`o3`, `o3-pro`)	`gpt-5.5` / `gpt-5.4-mini` / `gpt-5.4-nano` / `gpt-5.5-pro`

A few patterns stand out. The original Codex had the shortest life of any OpenAI model line: it was deprecated on March 20, 2023 and shut down just three days later, on March 23, 2023. The entire GPT-3 generation - base models, InstructGPT text models, and first-generation embeddings - retired together on January 4, 2024. GPT-4.5 lasted only about three months in the API (deprecated April 14, 2025, shut down July 14, 2025) before being folded back into the GPT-4.1 and GPT-5 lines. The pace has accelerated in the GPT-5 era: within roughly a year of GPT-5's launch, OpenAI scheduled retirements of the DALL·E image models (May 12, 2026), the Sora 2 video models and Videos API (September 24, 2026), a large wave of legacy chat models including gpt-3.5-turbo and the remaining o-series models (October 23, 2026), and the original GPT-5 and o3 snapshots themselves (December 11, 2026).

Separately from the API lifecycle above, OpenAI also retires models from the ChatGPT product itself: on February 13, 2026, GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini were retired from ChatGPT (with Custom GPT access for business plans ending April 3, 2026), moving remaining usage to the GPT-5 generation. Source: Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT. Always check the official deprecations page for the current schedule.

Frequently Asked Questions

When did ChatGPT launch?

ChatGPT launched on November 30, 2022 as a free research preview, built on the GPT-3.5 series. The paid ChatGPT Plus tier followed in February 2023. See the official Introducing ChatGPT announcement.

When was GPT-4 released?

GPT-4 was announced on March 14, 2023 as a large multimodal model accepting text and image input, with 8K- and 32K-token context variants. See GPT-4.

What is the difference between GPT-4 and GPT-4o?

GPT-4 (March 2023) accepted text and image input and produced text output. GPT-4o ("omni", May 2024) was designed from the start as a single model handling text, audio, and image input and output, with lower latency and a 128K-token context window, and a smaller GPT-4o mini followed in July 2024. See Hello GPT-4o.

When did the o1, o3, and o4-mini reasoning models come out?

The o-series reasoning models began with o1-preview and o1-mini on September 12, 2024, followed by the full o1 in December 2024, o3-mini on January 31, 2025, and o3 and o4-mini on April 16, 2025. These models are trained to produce an internal chain of thought before answering. See Introducing OpenAI o1-preview and Introducing OpenAI o3 and o4-mini.

What is the GPT-5 generation, and what is OpenAI's current model?

GPT-5 was announced on August 7, 2025 as a unified system that pairs a fast model with a deeper "GPT-5 thinking" reasoning model behind a real-time router. The generation advanced through GPT-5.1 (November 2025), GPT-5.2 (December 2025), GPT-5.4 (March 2026), and GPT-5.5 (April 23, 2026, with GPT-5.5 Instant becoming the ChatGPT default on May 5, 2026). GPT-5.6, released publicly on July 9, 2026 after a limited preview that began on June 26, 2026, is the current frontier generation as of the time of writing, offered in the Sol, Terra, and Luna capability tiers. See Introducing GPT-5 and GPT-5.6.

What do Sol, Terra, and Luna mean in GPT-5.6?

With GPT-5.6 (July 9, 2026), OpenAI replaced the Instant / Thinking / mini / nano naming used through GPT-5.5 with named capability tiers: the number (5.6) identifies the model generation, while GPT-5.6 Sol (the frontier tier), GPT-5.6 Terra (a balance of intelligence and cost), and GPT-5.6 Luna (the fastest, most cost-efficient tier) identify durable capability tiers that can advance on their own cadence. See GPT-5.6.

What happened to the o-series reasoning models in the GPT-5 era?

With GPT-5, the separate o-series brand was effectively folded into the unified GPT-5 system: the fast path corresponds to "Instant" models and the reasoning path to "Thinking" models, selected automatically by the router. There is no "o5"; reasoning is now a mode of the GPT-5 generation rather than a separate model family.

What is the difference between the original Codex and the new Codex?

The original Codex (announced August 10, 2021) was a GPT-3 descendant fine-tuned on code; it powered the first GitHub Copilot and was shut down in March 2023. The new Codex (2025) is an unrelated coding agent - first the open-source Codex CLI (April 2025), then a cloud software-engineering agent in ChatGPT (May 2025) powered by codex-1 - that reached general availability in October 2025 and initially ran on dedicated GPT-5-Codex models before the mainline GPT-5.5 and GPT-5.6 models began shipping directly in Codex. In July 2026, the Codex app was merged into the new ChatGPT desktop app. See Introducing Codex.

When did GPT models become available on Azure and Microsoft Foundry?

Microsoft announced the Azure OpenAI Service in November 2021 and made it generally available on January 16, 2023. GPT-4 reached Azure in preview on March 21, 2023, and recent models such as GPT-4o (May 2024) and the GPT-5 family (August 2025) have arrived on Azure at or near their OpenAI launch. The surrounding development platform was renamed from Azure AI Studio to Azure AI Foundry in November 2024, and to Microsoft Foundry in November 2025. See the Azure OpenAI "What's new" page.

What powered GitHub Copilot, and when did it launch?

GitHub Copilot launched in technical preview on June 29, 2021, powered by the original OpenAI Codex model, and reached general availability on June 21, 2022. It has since moved to newer GPT and Codex models. See Introducing GitHub Copilot.

When were the older GPT models such as GPT-3.5, GPT-4-32k, and GPT-4.5 deprecated?

See the Deprecations and Model Lifecycle section above for the full table. In summary, the GPT-3 generation retired on January 4, 2024, gpt-4-32k shut down on June 6, 2025, gpt-4.5-preview shut down on July 14, 2025, chatgpt-4o-latest shut down on February 17, 2026, the DALL·E models shut down on May 12, 2026, and gpt-3.5-turbo and the remaining o-series models are scheduled to shut down on October 23, 2026. In the ChatGPT product, GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini were retired on February 13, 2026. The most authoritative source is the official OpenAI deprecations page.

When did OpenAI add vision, voice, and image generation?

Vision (image input) arrived with GPT-4 in March 2023; voice and image input came to ChatGPT in September 2023; DALL·E provided text-to-image generation from January 2021 (DALL·E 3 in ChatGPT from October 2023), and GPT-4o native image generation replaced DALL·E 3 as the default image generator in March 2025, succeeded in turn by gpt-image-2 (ChatGPT Images 2.0) on April 21, 2026. Whisper added open-source speech recognition in September 2022, and the Realtime API and Advanced Voice brought speech-to-speech conversation from October 2024, with the full-duplex GPT-Live-1 models replacing Advanced Voice Mode in ChatGPT from July 2026.

Summary

In this article, I built a historical timeline of OpenAI's GPT models from GPT-1 in 2018 through the GPT-5 generation in 2025 and 2026, organized by the model family tree, the major release timeline, the evolution of capabilities, the products and services built on GPT, availability across platforms, and the deprecation lifecycle.

In a few years, OpenAI has gone from a single research model to a unified GPT-5 system, by way of the base GPT line, the multimodal GPT-4o models, and the o-series reasoning models that converged into GPT-5 - while building ChatGPT, Codex, DALL·E, Sora, Whisper, and the Assistants, Responses, and Realtime APIs on top of those models, and making them available through the OpenAI API, Azure OpenAI in Microsoft Foundry, Amazon Bedrock, Microsoft Copilot, and GitHub Copilot. The current frontier generation as of the time of writing is GPT-5.6, released publicly on July 9, 2026 in the Sol, Terra, and Luna capability tiers.

I will continue to monitor how OpenAI's release cadence and capability roadmap evolve. The most reliable place to track new model announcements is the official OpenAI News page, with model deprecations listed on the official OpenAI deprecations page.

In addition, there are also related model-timeline and history articles on hidekazu-konishi.com, so please have a look if you are interested.

Anthropic Claude Model Release Timeline - Model Family Tree, Capability Evolution, and Platform Availability
AWS Generative AI History and Timeline - From SageMaker JumpStart to Bedrock AgentCore
Amazon Bedrock Model Catalog 2026
AWS History and Timeline - Almost All AWS Services List, Announcements, General Availability(GA)

References

References:
Tech Blog with curated related content

Written by Hidekazu Konishi