Codex Timeline: Release Dates and Major Updates (2026)

Codex has had two distinct lives at OpenAI. The first began on August 10, 2021, when OpenAI opened a private beta for a model that translated plain-English instructions into code. That early Codex helped power GitHub Copilot and introduced many developers to AI-assisted programming before OpenAI retired the original model family in March 2023.

OpenAI brought the name back two years later for a different kind of product. Codex CLI arrived on April 16, 2025, followed one month later by a cloud agent that could enter a repository, change files, run commands and tests, and prepare the result for review. Codex had moved from generating code in response to a prompt to carrying a software task through several steps.

The releases that followed changed where Codex could work and how long a task could run. Editor support and general availability came first. Desktop and mobile apps followed, along with remote work, automation, and a succession of coding-focused models. The timeline below follows that history from the 2021 model through the modern Codex agent.

Last updated: July 21, 2026.

Latest Codex update

On July 9, 2026, OpenAI released GPT-5.6 across Codex and moved Codex into the ChatGPT desktop app on macOS and Windows. GPT-5.6 introduced three model tiers named Sol, Terra, and Luna, with availability varying by ChatGPT plan. The desktop integration preserved existing Codex projects, settings, and workflows while placing them inside the main ChatGPT application.

Codex timeline at a glance

Date	Event	What Changed
July 9, 2026	ChatGPT desktop integration	Codex joined the ChatGPT desktop app on macOS and Windows.
July 9, 2026	GPT-5.6	Sol, Terra, and Luna became available in Codex.
June 25, 2026	Codex Remote GA	Paired mobile devices could control work on Mac and Windows hosts.
June 2, 2026	Sites and role-specific plugins	Codex added hosted Sites and plugin collections for specific professions.
May 26, 2026	Model deprecations	GPT-5.3-Codex and GPT-5.2 left the ChatGPT-authenticated model picker.
May 14, 2026	Mobile preview and Remote SSH GA	Codex entered ChatGPT mobile, and remote SSH access reached GA.
May 7, 2026	Codex for Chrome	A browser extension added controlled access to signed-in tabs.
April 23, 2026	GPT-5.5 and browser use	A new default model and active in-app browser operation arrived.
April 16, 2026	Expanded app workspace	Chats, computer use, automations, artifact previews, and memories expanded the app.
March 25, 2026	Plugins	Installable bundles combined Skills, app integrations, and MCP configuration.
March 6, 2026	Codex Security preview	Aardvark became a repository-aware application security agent in Codex.
March 5, 2026	GPT-5.4	A general-purpose frontier model with native computer use entered Codex.
March 4, 2026	Windows app	The native app arrived with PowerShell and Windows sandbox support.
February 12, 2026	GPT-5.3-Codex-Spark preview	A low-latency model targeted real-time coding.
February 5, 2026	GPT-5.3-Codex	The model added stronger general reasoning and mid-turn steering.
February 2, 2026	Codex app for macOS	A desktop interface organized parallel agents, worktrees, and reviews.
January 14, 2026	GPT-5.2-Codex API	API-key workflows gained access to the model.
December 19, 2025	Agent Skills	Reusable instruction packages arrived in the CLI and IDE extension.
December 18, 2025	GPT-5.2-Codex	Long-horizon, Windows, vision, and security performance improved.
November 19, 2025	GPT-5.1-Codex-Max	Multi-context compaction supported longer agent tasks.
November 13, 2025	GPT-5.1-Codex models	Standard and Mini variants expanded the model lineup.
October 6, 2025	General availability	Codex reached GA with Slack, SDK, admin, and GitHub tools.
September 15, 2025	GPT-5-Codex and IDE extension	Codex gained a coding model, rebuilt CLI, and editor interface.
June 3, 2025	Plus expansion	ChatGPT Plus access and task internet controls arrived.
May 16, 2025	Cloud research preview	Codex began running delegated engineering tasks in isolated environments.
April 16, 2025	Codex CLI	OpenAI released its open-source local coding agent.
March 2023	Original models deprecated	The 2021 Codex API model family was retired.
August 10, 2021	OpenAI Codex private beta	The natural-language-to-code model entered API testing.

2021: OpenAI Codex begins as a coding model

August 10, 2021: Codex enters API private beta

OpenAI introduced OpenAI Codex as a system that translated natural-language instructions into code. It descended from GPT-3 and had training exposure to natural language and publicly available source code. OpenAI described Python as its strongest language, with support for more than a dozen others.

The first release was an API model, not an autonomous coding agent. Developers supplied prompts and built applications around the output. GitHub Copilot, launched through an OpenAI and GitHub partnership one month earlier, used Codex as its underlying model. The private beta also showed a wider aim: natural-language commands could call software APIs, which connected code generation with actions inside other programs.

2023: the original Codex models are retired

March 2023: OpenAI deprecates the Codex API models

OpenAI retired the original Codex model family in March 2023. Its historical announcement identifies the month but does not publish a specific retirement day. The timeline therefore keeps the official month-level date. Code generation continued through newer general models, but the Codex name disappeared from the active model catalog for two years.

2025: Codex returns as a coding agent

April 16, 2025: Codex CLI is released

OpenAI revived the name with Codex CLI, an open-source agent that ran in a terminal. It could inspect a repository, edit files, and execute commands in the developer’s own environment. This product behaved differently from the 2021 API model because the agent managed a work loop around the model: gather context, propose or apply changes, run tools, inspect results, and continue.

The CLI also established the local side of the later Codex product. Developers could keep their project, dependencies, shell tools, and Git state on their computer. Approval and sandbox settings governed which actions the agent could take. OpenAI published the source repository, where anyone could inspect releases and implementation changes outside the hosted product.

May 16, 2025: Codex cloud enters research preview

The Codex research preview added a cloud agent inside ChatGPT. A user could assign several software engineering tasks, and each task ran in a separate isolated environment prepared with the selected repository. Codex could write features, answer questions about a codebase, fix bugs, run tests, and prepare changes for review.

The cloud agent used codex-1, a version of o3 optimized for software engineering. OpenAI also released codex-mini-latest, an o4-mini variant for low-latency work in Codex CLI and the Responses API. ChatGPT Pro, Business, and Enterprise users received the first cloud access. On June 3, OpenAI extended access to ChatGPT Plus and opened configurable internet access for tasks that needed dependencies or external resources.

September 15, 2025: GPT-5-Codex and the IDE extension arrive

GPT-5-Codex was the first GPT-5 model optimized specifically for agentic software engineering in Codex. OpenAI trained it for short interactive work and longer independent tasks. It became the default for cloud tasks and code review, while local users could select it in the CLI and new IDE extension.

The same release rebuilt the CLI around longer agent workflows. Images, task lists, web search, MCP connections, clearer approval modes, and conversation compaction became part of local work. The IDE extension brought Codex into VS Code, Cursor, and other VS Code forks. On September 23, API-key users also received GPT-5-Codex through the Responses API.

October 6, 2025: Codex reaches general availability

OpenAI declared Codex generally available after its cloud research preview. The product connected terminal, editor, cloud, GitHub, and ChatGPT account workflows. General availability also brought a Slack integration, a TypeScript Codex SDK, workspace administration controls, and a GitHub Action.

The SDK exposed the same agent used by Codex CLI for custom tools and internal workflows. Slack messages could send a task to Codex cloud, while the GitHub Action supported automation inside CI. These additions changed Codex from a set of individual interfaces into a product that teams could place inside existing development systems.

November 2025: the GPT-5.1-Codex family expands

GPT-5.1-Codex and GPT-5.1-Codex-Mini joined the CLI and IDE extension on November 13. The standard model targeted long-running agentic coding, while Mini offered a lower-cost option with more usage under subscription limits. Six days later, GPT-5.1-Codex-Max became the default across Codex surfaces.

Max was the first Codex model trained to work across multiple context windows through compaction. That design supported project-scale refactors, extended debugging, and multi-hour agent loops without treating one context window as the hard boundary of a task. OpenAI also described it as the first model in this line trained for Windows environments.

December 18 and 19, 2025: GPT-5.2-Codex and Agent Skills

GPT-5.2-Codex improved long-context understanding, tool calls, large refactors, migrations, Windows work, vision, and defensive security tasks. It launched across Codex surfaces for paid ChatGPT users on December 18. API access followed on January 14, 2026.

Agent Skills arrived one day after the model release. A Skill packages reusable instructions with optional scripts and resources. Teams can keep these repeatable processes outside one conversation. Initial support covered Codex CLI and the IDE extension. Skills later became a foundation for desktop workflows and plugins.

2026: Codex becomes a multi-surface work agent

February 2, 2026: the Codex app launches on macOS

The Codex app gave local agents a dedicated desktop interface. Projects could hold several threads, and built-in worktree support isolated concurrent changes to the same repository. The app also included review tools, Git controls, Skills, Automations, and shared history with the CLI and IDE extension.

This release made parallel work visible without forcing every task into a terminal window. A developer could switch among projects, inspect diffs, comment on changes, and open files in an editor. The app launched on macOS for paid ChatGPT plans, with temporary access for Free and Go users.

February 5 and 12, 2026: GPT-5.3-Codex and Spark

GPT-5.3-Codex combined coding performance with stronger professional reasoning and research. It also responded to steering while a turn was running, which reduced the need to stop a long task and restart with a revised prompt. The model launched in the app, CLI, IDE extension, and cloud for paid ChatGPT plans.

GPT-5.3-Codex-Spark followed as a research preview for Pro users. It was a smaller model served on low-latency hardware for real-time coding. The pair separated two working styles: persistent execution with the main model and fast interaction with Spark.

March 4 and 5, 2026: Windows and GPT-5.4

The Codex app reached Windows on March 4. It ran natively with PowerShell and a Windows sandbox, while users who preferred Linux tools could select WSL. The app carried over parallel threads, worktrees, review, Skills, and Automations from the macOS edition.

On March 5, GPT-5.4 entered Codex. It was a general-purpose frontier model that incorporated the coding capability of GPT-5.3-Codex and extended it across tools, documents, spreadsheets, presentations, and computer use. Codex also offered experimental support for a one-million-token context window with this model.

March 6, 2026: Codex Security enters research preview

OpenAI renamed its Aardvark security agent as Codex Security and opened a research preview through Codex web. It builds a threat model for a repository, investigates possible vulnerabilities, validates findings in isolated environments where possible, and proposes patches for human review.

Codex Security had separate access rules from ordinary coding sessions. The preview rolled out to Pro, Business, Edu, and Enterprise users. Its inclusion under the Codex name added a specialized application security surface, but the product did not replace code review, tests, or a team’s security process.

March 25, 2026: plugins package tools and workflows

Codex plugins joined the app, CLI, and IDE extension. Each installable package could contain Skills, app integrations, and MCP server configuration. A plugin therefore handled a wider extension boundary than a Skill alone. Teams could install curated packages or develop local plugins and marketplaces for a repository or user account.

April 16 and 23, 2026: computer use and projectless chats

A major app update on April 16 opened chats that did not require a project folder. It also added an early in-app browser, computer use on macOS, scheduled thread Automations, artifact previews, memories, remote connection work, and fuller pull request review. Codex could now start from a document, a browser task, or a research question as well as a code repository.

GPT-5.5 became the recommended model for most Codex tasks on April 23. The bundled browser plugin also gained active operation of local development pages. Codex could click through an interface, reproduce a visual defect, and check a local fix inside the app. Automatic approval review entered the product at the same time for eligible actions.

May 7 and 14, 2026: Chrome, mobile, and Remote SSH

Codex for Chrome arrived on May 7. The extension could work with browser pages that required an existing signed-in session, and it could operate across tabs in the background. Users controlled which sites were available to the agent.

One week later, Codex entered the ChatGPT mobile app in preview on iOS and Android. A phone could start work, inspect output, approve commands, change models, and steer threads running on a connected machine. Remote SSH and Hooks reached general availability. The files, credentials, and execution environment stayed on the host while the mobile app carried the live task state.

May 26, 2026: older selectable models are deprecated

GPT-5.3-Codex and GPT-5.2 left the model picker for Codex sessions authenticated through ChatGPT. OpenAI directed users toward current models such as GPT-5.5, GPT-5.4, and GPT-5.4 mini. This deprecation applied only to the ChatGPT authentication path. API-key workflows were unaffected.

June 2 and 25, 2026: Sites and Codex Remote

OpenAI introduced Sites, role-specific plugins, and annotations on June 2. Sites could create and host websites, dashboards, internal tools, web apps, and games from the Codex app. Initial Sites access was a preview for Business and Enterprise workspaces. The new plugins targeted work beyond software engineering, while annotations supported precise revision of code and other artifacts.

Codex Remote reached general availability on June 25. Authenticated one-to-one QR pairing connected each phone to a Mac or Windows host. The mobile app could then start or continue tasks, display progress, and handle approvals. Remote work now had a supported general-availability path after the earlier mobile preview.

July 9, 2026: GPT-5.6 and ChatGPT desktop integration

GPT-5.6 launched in Codex with three model tiers. Sol targeted the hardest work, Terra balanced capability and cost, and Luna focused on speed and lower cost. Free and Go users received Terra, while paid plans included additional model options. Codex also supported the max effort setting, with ultra available on eligible paid plans.

Codex joined the ChatGPT desktop app on the same date. Users of the separate Codex app could update normally and keep their projects, settings, and workflows. The release added direct editing, inline annotations, pull request chat, custom domains for Sites, and faster computer use with GPT-5.6. This integration tied Codex more closely to the main ChatGPT application while retaining its project and agent workflow.

Where Codex runs

Codex CLI runs in a local terminal and works with files, commands, Git, MCP servers, scripts, and non-interactive automation.
The IDE extension places local sessions, diffs, and cloud task handoff inside VS Code, Cursor, and compatible editors.
Codex cloud runs delegated tasks in managed environments connected to a repository and configured setup process.
The ChatGPT desktop app manages projects, parallel threads, worktrees, reviews, terminals, Automations, browser tools, computer use, and plugins.
The ChatGPT mobile app can start and supervise Codex work on connected local or remote hosts.
The SDK, GitHub Action, Slack integration, hooks, access tokens, and non-interactive CLI connect Codex with team systems and automation.

Models that changed Codex

The modern cloud agent began with codex-1, while codex-mini-latest served lower-latency CLI and API work. GPT-5-Codex then aligned the model with the agent’s complete work loop: reading a repository, editing files, running tools, reviewing code, and continuing through test failures.

The GPT-5.1-Codex and GPT-5.2-Codex releases concentrated on longer tasks, context compaction, Windows, vision, and security work. GPT-5.3-Codex combined coding with reasoning across research and professional tasks, while Spark tested a faster interactive path. Model deprecations in May 2026 showed that model availability inside a ChatGPT-authenticated Codex session could change faster than the product interfaces.

GPT-5.4 changed the naming pattern because a mainline general model became the recommended Codex choice. GPT-5.5 continued that direction. GPT-5.6 then introduced durable Sol, Terra, and Luna tiers across ChatGPT, Codex, and the API.

How Codex changed over time

The 2021 Codex model converted prompts into code for applications built through an API. The 2025 agent managed the surrounding work: it gathered project context, changed files, ran commands, tested results, and presented evidence for review. That agent loop became the stable connection between later models and interfaces.

Codex then spread across the places where work happens. The terminal and editor supported close supervision. Cloud tasks handled delegated work in managed environments. The desktop app organized concurrent projects and review, while mobile and remote access supported check-ins away from the host. SDK, Slack, GitHub, hooks, and plugins connected the agent to team systems.

The 2026 releases widened the type of work Codex could accept. Browser and computer use added visual interfaces. Artifact previews, Sites, annotations, plugins, and projectless chats supported documents, research, analysis, and hosted tools.

Codex FAQs

Is Codex free?

Codex access depends on the ChatGPT plan, product surface, and current promotion. OpenAI has offered limited Codex access to Free and Go users, while paid plans include higher usage limits. API-key use has separate usage billing.

Does Codex run code locally?

Local Codex sessions can read files and run approved commands on the selected computer or SSH host. Sandbox rules, writable folders, and approval settings limit those actions. Codex cloud uses an OpenAI-managed environment. A cloud task receives only the files, dependencies, credentials, and services configured for that environment.

Is Codex CLI open source?

Yes. OpenAI publishes Codex CLI at github.com/openai/codex. The repository contains the agent implementation, installation instructions, source changes, and numbered releases. Hosted models and managed Codex services are separate from the open-source CLI code.

Does Codex require GitHub?

No. Codex CLI and the desktop app can work with local folders and Git repositories without a GitHub connection. GitHub is important for cloud repository tasks, pull requests, code review, and the GitHub Action. A local session can also use another Git host through ordinary command-line tools if the environment has access.

Is there a Codex API?

Developers can automate Codex through the Codex SDK, non-interactive CLI commands, GitHub Action, hooks, and supported access tokens. Some Codex-related models have also been available through the Responses API. Model API access and the Codex agent product are different layers. A model endpoint does not reproduce the complete CLI or cloud agent workflow by itself.

Can Codex use browser pages and desktop applications?

Supported Codex app versions can operate an in-app browser and use desktop applications through computer use. The Chrome extension can supply signed-in browser context for approved sites. Availability depends on the operating system, region, plan, administrator controls, installed plugin, and permission settings.

How should I choose a Codex model?

The model recommended in the current Codex picker is the default choice for most work. Faster or lower-cost tiers suit short edits and frequent interaction, while higher-capability tiers handle difficult debugging, large refactors, research, and long agent runs.