OpenAI Codex

OpenAI Codex is a family of AI software-engineering products developed by OpenAI. The name has covered two distinct things: a 2021 code-generation model, a GPT-3 derivative that powered the first GitHub Copilot and whose standalone API was retired in 2023, and a 2025 agent that reads a repository, edits files, runs commands and tests, and proposes or commits changes. The current product comes in several surfaces, including an open-source terminal agent (artificial intelligence-based command-line tool), a cloud agent that runs each task in an isolated sandbox, and IDE and desktop integrations. These are built on OpenAI's reasoning and coding models, which were trained with reinforcement learning on software tasks.

The product matters because it is a central example of the shift from code autocompletion and chat assistants toward AI agents that carry out multi-step work with limited supervision. It competes with other agentic coding tools, and its expansion beyond programming toward general computer-use tasks raises questions about what a coding agent is and how much autonomy it should have. The entry first separates the two generations that share the name, then traces the 2025 relaunch, the architecture and model lineage, and the expansion into new surfaces and access tiers. It then covers engineering workflows, adoption, safety and limitations, the open-source ecosystem, and unresolved debates.

Name and scope of the term

The word "Codex" now covers several related but distinct things. The oldest is a 2021 model. It was a GPT-3 variant trained further on public source code, and it turned plain-English descriptions into programs. It served as the original model for GitHub Copilot. That model was a code-generation and autocomplete engine. It was not an agent, and it is not the product that carries the name today.

The second thing is the 2025 Codex agent family, which began as a terminal tool and then gained a hosted counterpart. The third is a desktop application that, by 2026, has been extended well beyond software engineering. The three share a brand, and some share underlying models, but they differ in purpose, licensing and deployment.

Between the two uses of the name lies a gap of roughly two years. OpenAI deprecated and shut the original model down in March 2023, advising developers that its newer general-purpose GPT models were better at coding. The deprecation notice reportedly said that "our latest models are now our best models for coding-related tasks". Copilot moved to GPT-4 in the same period, and the name "Codex" went on hiatus while code-focused LLMs evolved. What was discontinued was the standalone code model and its API. Its function was absorbed into models that were not specialized for code alone. The name returned in April 2025 attached to an agent product, not a model.

The current products are separate. OpenAI released Codex CLI on April 16, 2025, as an open-source coding agent that runs locally in the terminal. A month later, on May 16, 2025, the company announced a research preview of Codex Cloud. The cloud service runs tasks on remote infrastructure rather than on the developer's machine. The desktop app is a third surface. Its licence is proprietary, while the CLI is licensed under Apache 2.0 (open-source software licensing is covered in a separate section). Reports also differ on timing. One account says the launch on February 2, 2026 was the macOS launch, while another says the Windows expansion came as of early March 2026. A third source says Windows support arrived on May 29, 2026. A blog post also claims that the desktop app was merged into the ChatGPT desktop app in July 2026. That claim rests on a single secondary blog and should be checked against OpenAI's own release notes.

The term "coding agent" marks a difference in kind, not just scale. Code completion works inside an editor and predicts the next few tokens or lines as the developer types. Earlier tools such as GitHub Copilot focused on autocompleting lines of code in real time. Chat-based generation takes a prompt and returns a snippet, which the developer must still copy, integrate, run and debug. A coding agent takes a task and carries it through several steps. It reads the repository, edits files, runs commands, and checks the outcome. It reads the code, proposes and applies edits, runs commands inside a sandbox with approval controls, and keeps a resumable session history. The difference is therefore one of autonomy and scope: the output is an observable change set to be reviewed, not a suggestion to be accepted or rejected line by line.

The original Codex model (2021–2023)

OpenAI announced the first Codex in August 2021. It was a modified production version of GPT-3, further trained on public source code in several programming languages, and it was built to read plain-language instructions and produce working code. It served as the original model for GitHub Copilot. One widely cited training detail is a filtered set of 159 gigabytes of Python code sourced from 54 million public GitHub repositories. The model was offered through an API in a limited private beta, which is why API access was the main route for outside developers.

Its most visible use was inside GitHub Copilot. GitHub introduced that tool in mid-2021 as a technical preview built on Codex. In practice, it worked as an autocomplete system: it analyzed the lines a programmer had already written to suggest new code, and even whole functions, which the developer could accept or ignore. Suggestions were strongest in mainstream languages, and early documentation named Python, JavaScript, TypeScript, Ruby, and Go as the best-supported ones. This is the sense of "code completion" that the later agent products moved beyond. Critics at the time also pointed out that the training code had not been validated for correctness or security, so the output carried the same kinds of errors that the underlying language model produced.

The standalone model was retired abruptly. OpenAI announced the Codex deprecation on 20 March 2023 and switched the models off on 23 March 2023. The reason given to customers was that newer models had become better at coding tasks. Developers were told to migrate to GPT-3.5-Turbo or GPT-4 instead. In effect, the coding capability was folded into general-purpose large language models rather than kept as a separate specialist line. Copilot followed the same path. GitHub later said that Copilot has transitioned away from Codex and now defaults to the latest frontier models, and Copilot Chat moved to GPT-4 in November 2023. Secondary commentary gives different accounts of exactly when Copilot changed models in early 2023, so that sequence should be checked against GitHub's own changelog.

The lifecycle shows several things about how the company handled code-specific models. First, the code model was treated as a stepping stone, not a permanent platform. General models that could also write code overtook it in less than two years. Second, the product was offered as a low-level primitive, a text-in, code-out API, rather than as a finished tool, and it was shut down with only days of notice. Third, although it was labeled a beta, some teams had already put production traffic on it, which made the shutdown costly for them. Fourth, the name then lay dormant for about two years before being reused for a different kind of product, the agent described in the next section. One secondary source reports that research access to the API was later restored after researchers objected. That claim, along with its dates, could not be confirmed from primary sources and should be treated with caution.

Relaunch as an agent (2025)

The 2025 relaunch began with a local tool. OpenAI released Codex CLI on April 16, 2025, as an open-source coding agent that runs in the terminal and connects its language models to local code, shell commands and files (OpenAI released Codex CLI on April 16, 2025, as an open-source coding agent that runs locally in the terminal. It connects OpenAI's language models with local code and command-line tasks). Because it works inside the developer's own environment rather than on a hosted copy of the project, it can use the toolchain, dependencies and files the developer already has. OpenAI's own update log also dates the CLI launch to April 2025 (Update on April 16, 2025: We launched Codex CLI, our new open source local coding agent.). Some third-party guides instead give May 2025 as the CLI's date (OpenAI Codex CLI is OpenAI's open-source, command-line coding agent, first released in May 2025), which appears to conflate it with the cloud product announced that month.

Cloud research preview (May 2025)

A month later, on May 16, 2025, OpenAI announced a research preview of a cloud-hosted agent (A month later, on May 16, 2025, the company announced a research preview of Codex Cloud.). Each task ran in its own cloud sandbox, started from a copy of the user's repository, and could cover feature work, questions about the codebase, bug fixes and pull-request proposals (Codex can perform tasks for you such as writing features, answering questions about your codebase, fixing bugs, and proposing pull requests for review; each task runs in its own cloud sandbox environment, preloaded with your repository.). Multiple tasks could be started at once, which distinguished the product from single-prompt tools.

The model behind the preview was codex-1, a variant of OpenAI's o3 reasoning model tuned for software engineering (Codex is powered by codex-1, a version of OpenAI o3 optimized for software engineering.). According to OpenAI, it was trained with reinforcement learning on coding tasks, with the training favoring repeated test runs until the results passed (The company said the model was trained using reinforcement learning on real-world coding tasks and could iteratively run tests until it received passing results.). This training target is the basis for the test-driven behavior described in later sections.

Access was staged. The preview first reached the higher paid tiers, with Plus promised later (Available to ChatGPT Pro, Team, and Enterprise users today, and Plus users soon.). Plus users gained access in June 2025 (The research preview first rolled out to some paid ChatGPT tiers, and ChatGPT Plus users gained access in June 2025.). Early cloud tasks ran without internet access; a June update allowed controlled access so the agent could fetch documentation and install packages (It was then integrated into ChatGPT as a sandboxed cloud agent in May, operating initially without internet access. A pivotal update in June granted Codex controlled internet access, allowing it to fetch documentation and install software packages.).

General availability and team features (October 2025)

OpenAI announced general availability on October 6, 2025, alongside three additions aimed at engineering teams: a Slack integration, an SDK and admin tools (OpenAI Codex is now generally available with powerful new features for developers: a Slack integration, Codex SDK, and admin tools like usage dashboards and workspace management). In Slack, a user tags the agent in a channel or thread; it gathers context from the conversation, picks an environment and replies with a link to the finished task, which can then be merged, extended or pulled onto the user's own machine (Tag @Codex in a Slack channel or thread and it will automatically gather the context it needs from the conversation, choose the right environment, and answer with a link to the completed task in Codex cloud. From there, you can merge its changes, keep iterating, or pull the task to your computer to keep working locally.).

The SDK exposes the same agent implementation that drives the CLI for use in other tools and workflows, initially in TypeScript (With the Codex SDK, you can embed the same agent implementation that powers the Codex CLI in your own workflows and apps. Initially available in TypeScript with more languages coming.). The admin tools added environment controls, monitoring and analytics for workspace administrators (With environment controls, monitoring, and analytics dashboards, ChatGPT workspace admins now have more visibility and control to manage Codex at scale.). Slack and the SDK were offered across the paid plans, while the admin tools were limited to the business and enterprise tiers (The Slack integration and Codex SDK are available to developers on ChatGPT Plus, Pro, Business, Edu, and Enterprise plans starting today, while the new admin features will be available to Business, Edu, and Enterprise.).

From code generation to delegated execution

The 2021 model and the first Copilot integrations were built around producing or completing code in response to a prompt, with the developer checking each output and deciding the next step. The 2025 product changed the unit of work. The user assigns a goal, and the agent reads the repository, edits files, runs commands and tests, responds to failures, and returns a diff with its test results and a summary for human review (When the work is done, Codex packages a diff, the test results, and a summary into a pull request for human review.). Launch coverage framed this against autocomplete-style assistants, describing an agent that could carry out work without step-by-step prompting (Unlike traditional code assistants operating as autocomplete tools, Codex is a collaborative partner who can independently take on tasks.). One analysis described the cloud mode as the clearest example of the move from assistance to delegation (This is the surface that most clearly embodies the shift from assistance to delegation.). That description comes from a secondary commentary rather than from OpenAI.

Architecture and execution model

The agent loop

Codex works in a cycle rather than producing a single answer. A third-party explainer says the codex-1 model edits the code and runs your tests in a loop until they pass. OpenAI's own launch description lists the same kinds of actions: the agent can read and edit files, as well as run commands including test harnesses, linters, and type checkers. The loop therefore moves from reading the repository, to deciding what to change, to editing, to running checks, and back again when a check fails. The end product is a proposed set of changes. It returns a pull request with the full command log attached for review, before anything touches your main branch.

Cloud execution

In the cloud, each task gets its own workspace. Every job starts from a fresh sandbox seeded with the project's files, and the first step is to prepare that environment. Every cloud task, whether launched from the Codex app, an @codex mention on GitHub, or the CLI's cloud command, runs against a cloud environment that defines the repository, dependencies, secrets, and network policy. OpenAI's documentation says the default image is called universal and comes with common languages and tools preinstalled, and that Codex caches container state for up to 12 hours to speed up new chats and follow-ups.

Network access was initially disabled. The launch announcement stated that internet access is disabled, limiting the agent's interaction solely to the code explicitly provided via GitHub repositories and pre-installed dependencies. Later configuration lets an administrator enable outbound access per environment, restricted by domain allowlists and permitted HTTP methods.

Because each run is independent, tasks can proceed without the developer present. A third-party account of the September 2026 relaunch says the new flow lets a developer prepare and publish a cloud environment once, after which tasks run in parallel while your laptop is closed. This relaunch is documented mainly in secondary reporting, so its details should be checked against OpenAI's current documentation.

The local CLI: permissions, sandboxing and implementation

The open-source CLI runs in the developer's terminal and operates on a local Git checkout. It is written in Rust, published under the Apache-2.0 licence, and works against a local Git repository: it reads files, proposes and applies edits, runs shell commands, and pauses for approval according to a sandbox policy you set.

Two separate settings control its reach. The sandbox mode determines what commands can technically do, and the approval policy determines when the agent stops to ask. The documented sandbox modes are read-only, workspace-write, and danger-full-access. In workspace-write mode, writes are allowed inside the workspace roots, and outbound network stays off unless sandbox_workspace_write.network_access is set to true. The most permissive mode, danger-full-access, applies only in a disposable environment such as an isolated CI runner or container. For enforcement, the CLI relies on the host operating system rather than its own code. Reported implementations use Seatbelt on macOS, Bubblewrap on Linux, and restricted tokens on Windows. In the workspace-write mode, the CLI also protects the .git directory from edits by default.

Approval policies range from untrusted, where Codex pauses for approval on anything it does not already consider safe, to on-request, where the agent asks when it decides it needs to, to never, under which Codex never pauses. Guides disagree on the default values and on which legacy options have been removed, so the official configuration reference should be consulted. Settings are kept in a config file, where sandbox_mode and approval_policy are set at the top level, with per-profile overrides under named profiles.

Extending reach beyond the repository

The CLI and cloud agent can use external tools through the Model Context Protocol (MCP). The Model Context Protocol is how AI agents talk to external tools, with a server publishing callable tools over JSON-RPC so any MCP-aware agent can connect, read the list, and invoke them. Developers can add MCP servers by hand in the configuration file or through a command-line helper.

Plugins package these pieces for distribution. One account states that Codex CLI v0.117.0 (released March 26, 2026) elevated plugins to a first-class workflow primitive. A plugin bundles skills, MCP servers, and app connectors, so one install wires up all three. Connectors give the agent authenticated access to outside services such as GitHub, Slack, or Linear. A separate report says that an April 2026 update shipped a Codex update with more than 90 new plugins, which shows that third-party reporting of plugin dates is not fully consistent. Plugins also raise a security question, since a connected tool can supply instructions the agent then follows, a concern discussed in the safety section.

Model lineage

From codex-1 to GPT-5-Codex

The first models offered with the 2025 product were codex-1, an o3-based model tuned for coding, and codex-mini-latest, a smaller derivative released for the CLI and API. Both were trained for software tasks on top of general large language model architectures, so they were specialized rather than separate systems.

The first GPT-5 variant for the product arrived in mid-September 2025. It was introduced on 15 September 2025 as a GPT-5 variant tuned for agentic coding in Codex. Cloud tasks and code review used it by default, while local workflows in the CLI or IDE extension could select it manually. A week later it reached the Responses API, which made it available to developers outside the product, in addition to Codex surfaces. OpenAI also drew a clear division of labor: the coding-tuned variant handled software work, and the base GPT-5 model was intended for other tasks.

On 19 November 2025 the company released GPT-5.1-Codex-Max, aimed at project-scale work. It is faster, more capable, and more token-efficient than GPT-5.1-Codex, using compaction to work coherently across multiple context windows.

GPT-5.2-Codex and GPT-5.3-Codex

OpenAI released GPT-5.2-Codex on December 18, 2025, describing it as the most advanced agentic coding model yet for complex, real-world software engineering. Its design extended the general model: it builds on GPT-5.2's general capabilities and extends them with deeper context handling, stronger agentic coding performance, and enhanced security awareness. Third-party listings date its API availability to 14 January 2026, when it was priced at $1.75 input and $14.00 output per 1M tokens, with a 400K context window.

GPT-5.3-Codex followed on 5 February 2026. OpenAI credits it as the first model to unify the earlier Codex and GPT-5 training pipelines, and reports it as ~25% faster than its predecessor. Availability reports conflict. DataCamp reported that it was available across all paid ChatGPT tiers in the app, via the CLI, via an IDE extension, and on the web. Wikipedia, by contrast, stated that as of February 2026, GPT-5.3-Codex is only available for ChatGPT Pro ($200/month) subscribers. API access was delayed: OpenAI said it was "working to safely enable API access soon." A smaller, text-only sibling, GPT-5.3-Codex-Spark, entered a research preview on 12 February 2026.

Later entries in the sequence include GPT-5.4, which the Codex changelog listed as OpenAI's most capable and efficient frontier model for professional work and the recommended choice for most Codex tasks from 5 March 2026. OpenAI's release notes also record the retirement of GPT-5.5 from ChatGPT, ChatGPT Work and Codex on 14 October 2026. A release-notes aggregator reports GPT-6.1 Sol as the default in Codex as of October 2026, but this should be confirmed against OpenAI's own notes.

Merging coding optimization with general reasoning

The early Codex-specific models were derived from general models and then specialized. From GPT-5.2-Codex onward, the emphasis shifted toward work that needs both code generation and broader reasoning. One analysis describes GPT-5.3-Codex as a successor to 5.2-Codex whose scope covers everything from writing code and debugging to updating Jira tickets and creating documentation. OpenAI framed the change as a move from code generation to a general-purpose coding agent that users can steer while it works.

Tuning for long-running tasks

Earlier releases were judged mainly on producing correct code in bounded tasks. Later releases were tuned for sessions that run for many steps and exceed a single context window. Compaction is the main mechanism: GPT-5.1-Codex-Max used it across multiple context windows, and GPT-5.2-Codex reportedly relied on stronger long-context understanding and native context compaction that preserve task state across extended sessions. Reliability of tool use was a parallel focus, since more reliable tool calling supports uninterrupted multi-step workflows. For GPT-5.3-Codex, the change was visible to users: they could ask questions, discuss an approach, and adjust direction mid-task rather than waiting for a final answer.

How model choice is exposed

In the CLI, IDE extension and desktop app, users pick a model and, in most surfaces, a reasoning-effort level, and the defaults shift as OpenAI ships new versions. Developers reach these models through the API. GPT-5.2-Codex, for example, accepts text and images as input and returns text. Third-party evaluators have run the models at high settings; one benchmark configuration for GPT-5.2-Codex lists Reasoning Effort: high, which is a benchmarking choice rather than a default. Enterprise customers see the models through partners as well. GitHub made GPT-5.3-Codex the base model for all Copilot Business and Copilot Enterprise organizations, and described it as its first long-term support (LTS) model.

Keeping this information current

Model names, retirements and default settings change quickly, and secondary sites often disagree with each other about dates and access tiers. OpenAI's release notes cover ChatGPT, API Platform, and Codex and are the authoritative source for the current model list.

Surfaces and product expansion (2025–2026)

IDE extension and desktop app

Codex reached developers through several surfaces besides the terminal and the cloud. OpenAI introduced a standalone Codex app for macOS in early 2026 and described it as a command center for AI coding, meaning an interface for running several AI agents at once on long-running work. A developer-community summary describes the macOS app as working alongside the Codex CLI and the IDE extension, so the three surfaces share one agent rather than forming separate products. The app's feature set included parallel work with changes isolated in Git worktrees, reusable skills that package a team's tools and conventions, and scheduled automations that run in the background. OpenAI added Windows support on March 4, 2026. At DevDay in September 2026, a code review view was added to the desktop app.

GitHub pull requests and code review

Codex can review pull requests on GitHub. A repository must first be connected to Codex, and automatic reviews are switched on in its settings; a reviewer can also mention the Codex account in a pull-request comment to request a review. Repository rules are written in an AGENTS.md file. By default the GitHub review flags only the most serious issues (P0 and P1 severity) to keep comments focused. A more in-depth Security Review was in research preview at the time of the documentation. At DevDay 2026, automated first-pass reviews were extended to GitHub pull requests and GitLab merge requests. Code review is distributed as a plugin in the ChatGPT sidebar, and the local /review command applies the same kind of review to uncommitted changes or branch diffs before a pull request exists.

Secondary reporting on the plugin notes a practical hazard: comments written in certain views are posted to the hosting platform immediately, not when the reviewer submits, and OpenAI's own guidance tells users to check the file and line references in drafts before posting.

Cloud environments and CLI refresh (DevDay, September 2026)

At DevDay on September 29, 2026, OpenAI announced reusable cloud development environments. Teams can share one configuration with approved settings and permissions, and developers can start remote tasks from a computer, a phone or the cloud without rebuilding the project setup each time. Commentary from a secondary source characterizes this as a move away from disposable per-task containers toward persistent development infrastructure. According to a recap table, the cloud feature is available on Plus, Pro, Business, Healthcare, Edu and Enterprise plans.

The same event brought a revised Codex CLI. Voice input is triggered with a /voice command so that instructions can be spoken. The /agents view lists running tasks and lets the user switch between them, while /fork opens a parallel session with the same context in a separate Git worktree. Other changes include session resumption, older history that loads as the user scrolls, and a usage view. The CLI updates were listed as available on all plans.

OpenAI also announced Codex Security Cloud, which scans repositories and new commits, removes duplicate findings and prepares fixes; this is covered in the safety section.

Toward non-coding work

The DevDay announcements extended agent capabilities beyond software development. The Agents API gained computer use, which allows applications to operate software through graphical interfaces, and the GPT-6.1 Sol model was made available in Codex as well as in the API and ChatGPT Work. Separately, a recap table lists "Dots", persistent agents with a cloud computer, a browser and connected apps, which were rolling out first to Pro and Business Premium users in eligible markets. These products are distinct from the Codex coding agent, but they share infrastructure and point toward agents that handle general office and desktop tasks alongside code.

Access, pricing and enterprise use

Plan tiers and gated features

Codex has no dedicated price. Users reach the agent through a ChatGPT plan, and the plan sets how much work they can run. OpenAI's pricing page places Codex inside the Free, Go, Plus, Pro, Business and Enterprise plans. Cloud tasks are gated more tightly. Paid tiers from Plus upward, together with Business, Enterprise, Healthcare and Education accounts, can use the cloud agent, though rollout and workspace settings can narrow access. Free and Go subscribers lack cloud task access. Seat type matters as well: view-only seats, including view-only Enterprise seats, cannot create cloud environments.

Price points differ by source. One pricing guide lists Free at $0 a month, Go at $8, Plus at $20, Pro at $100, $200 or $500, Business at $20 per user on annual billing ($25 monthly), and Enterprise or Edu on custom pricing. The Pro tier names also vary. Business is described inconsistently: one guide calls it pay-as-you-go per seat, while OpenAI's pricing text refers both to the $20 Business plan and to the $100 Business plan. The $500 tier and its faster variant appear only in secondary reports.

Usage limits and credits

Plus users work within rolling windows. The pricing page publishes ranges of local messages per model, for example 15 to 160 for GPT-6.1 Sol and 350 to 3,000 for GPT-6 Luna, with weekly limits that may also apply. Pro plans currently have no five-hour limit. Credits extend usage beyond these caps. Plus and Pro users can buy credits without switching plans. According to one billing guide, OpenAI moved from per-message billing to token-based credit billing for Plus, Pro, Business and new Enterprise plans on April 2, 2026. Under that model, GPT-5.3-Codex was listed at 43.75 credits per million input tokens and 350 per million output tokens. OpenAI publishes credit rates but not a fixed dollar price per credit, so the four-cent conversion that some third-party analyses use is an estimate.

Enterprise and Edu accounts on flexible pricing face no fixed rate limits, and their usage scales with purchased credits.

API access and per-token pricing

Codex-specific models can also be called directly. GPT-5-Codex is a GPT-5 variant tuned for agentic coding, is available only through the Responses API, and its underlying snapshot is updated regularly. The listed rates are $1.25 per million input tokens, $0.125 for cached input and $10 for output. A later model, GPT-5.3-Codex, was released on February 24, 2026, and an aggregator lists it at $1.75 per million input tokens and $14.00 per million output tokens. API billing sits outside the subscription. Using an API key bills per token and draws nothing from plan allowances. Developers who call the API directly do not receive cloud-only features such as GitHub code review and Slack integration, and new models reach API users later than ChatGPT subscribers.

Team administration and enterprise controls

Organizational controls are tiered. One guide lists Business as adding SAML SSO, MFA and no training on customer data, and Enterprise as adding SCIM, EKM, RBAC, audit logs and data residency. OpenAI's help documentation also states that Codex Enterprise Analytics is available to Enterprise workspaces with Codex enabled. Cloud availability depends on rollout and workspace settings, which administrators control. The sources reviewed here do not give detailed documentation of sandbox and approval policies at the workspace level, so those settings are described only in general terms.

Adoption alongside other coding agents

Cost comparisons place Codex close to its rivals. One analysis states that both Codex and Claude Code top out around $200 and can bill past their allowance, Claude Code through usage credits and Codex through purchasable credits since October 2025. Spending estimates are also company-issued. A blog reports that OpenAI's own help centre puts typical spend at roughly $100 to $200 per developer per month. Claims about organizations switching from other agents to Codex, or running it alongside them, rest mainly on vendor announcements and press relays. They should be read as company-reported rather than independently measured. Model names and rate tables in this area change frequently, so current figures should be checked against OpenAI's pricing page and release notes.

Engineering workflows and reception

Common uses

Developers have used Codex for writing features, fixing bugs, answering questions about an unfamiliar codebase, refactoring older code, and handing off routine chores. Early hands-on reports suggest it is most dependable for small, well-scoped work. One reviewer said it had been perfect for firing off maintenance-level updates such as minor copy tweaks, style changes and other small chores. The same reviewer found that larger refactors were cumbersome, partly because the workflow opened a fresh pull request for every iteration, so pushing follow-up commits to an existing branch was awkward.

Parallel tasks, worktrees and subagents

The Codex app lets a developer run several agents at once across projects. For isolated parallel work, it supports Git worktrees, with separate working directories linked to the same repository. Because worktrees share repository data rather than requiring separate clones, agents can edit in parallel without overwriting each other's files. The results still need human review: later merges still need review and may require conflict resolution. The app's diff view lets a reviewer comment on changes and then commit or reject them (review a given agent's output via a diff view, comment on changes, and commit or reject them).

Subagents, which were announced in general availability in March 2026 after a preview behind a feature flag, are specialized child agents that Codex spawns for parallel work. Simon Willison wrote that they are very similar to the Claude Code implementation, with default roles named explorer, worker and default. OpenAI's documentation recommends parallel agents for read-heavy tasks such as exploration, tests, triage, and summarization, and advises more caution for write-heavy work because agents editing code at once can create conflicts and increase coordination overhead. In practice, this shifts the developer's role from writing each change toward defining goals, supervising tasks, reviewing changes, and integrating results.

Reported productivity and criticism

Reactions are mixed and largely come from blogs, reviews and social media rather than controlled studies. Shortly after the release of GPT-5-Codex, one roundup reported that developers have reported meaningful productivity gains alongside some notable quirks. That roundup also relayed a claimed 90% drop in cloud task completion times, which comes from vendor-side infrastructure changes and has not been independently verified.

Critics have described different failure patterns. One engineering team's account found that Codex won't catch subtle logic bugs, and that the resulting burden shifts to the developer, who still has to write tests, scan edge cases, and debug regressions. The same account concluded that sometimes, that takes more time than if you'd just written it yourself. Another reviewer reported that starting tasks failed, and opening pull requests failed, and judged that for significant refactoring or feature work they were still better served doing that myself in an IDE with optional LLM support. Users on social media have also complained that usage limits on high-end models are depleting faster, which has prompted discussion of efficiency. Documented frustrations specific to data analysis and other non-software work were not found in the sources reviewed for this section.

Position among coding agents

Codex sits in a crowded field. Claude Code is a terminal-first agent from Anthropic, and comparison writers describe its default loop as running autonomously through plan → edit → test → fix → retry without intervention. Cursor is oriented around "propose a change, you approve, you continue." GitHub Copilot is widely described as a multi-editor extension, with the most lowest-friction option for teams already on GitHub. Cline is an open-source agent that runs inside VS Code, and it offers Plan and Act modes with a bring-your-own API key. Codex itself is also available through GitHub: GitHub added the agent in public preview to its Agent HQ system, making it available inside GitHub, GitHub Mobile, and Visual Studio Code for some Copilot subscribers. Many developers combine tools rather than choosing one; a common pattern pairs an IDE assistant for everyday editing with a terminal agent for larger refactors. Most head-to-head comparisons come from vendors, affiliates or independent blogs with small samples, so their rankings should be treated as indicative only.

Safety, reliability and limitations

The most characteristic failure of a coding agent is code that looks plausible but is wrong. The model may call a library function that does not exist, misremember a signature, or write logic that reads correctly and then mishandles an edge case. This is a software form of hallucination. The consequences are concrete: an invented import usually fails at build time, but a subtly wrong permission check can compile, pass a casual review and ship. For this reason, the agent's output is best treated as a draft. Human review catches problems that tests do not express, such as a misreading of the requirement or an unsafe design choice. Automated tests catch regressions a reviewer misses. The two do not substitute for each other. An agent that writes its own tests can encode the same misunderstanding in both the code and the test, so passing tests show consistency rather than correctness. Stronger guarantees, such as formal verification, are rarely applied to agent-written code in ordinary projects.

Sandboxing, network limits and approvals

Local Codex uses two independent controls. The sandbox is enforced by the operating system and defines what a command can technically reach, typically the current workspace. The approval policy determines when the agent must stop and ask the user before acting. The official documentation states that the agent runs with network access turned off by default. The CLI offers three sandbox levels: read-only, workspace-write and danger-full-access, the last of which removes sandbox restrictions entirely. The common "Auto" preset lets the agent read files, edit them and run commands inside the working directory without interruption, but asks before writing elsewhere or reaching the network. Enabling network access is an explicit configuration step, and some teams permit only named domains, such as package registries and source hosts. Administrators can lock these settings through managed configuration. Frequent approval prompts for the same network request are a signal that a narrower allowlist, rather than full access, is the better fix.

Prompt injection is a separate risk. An agent that reads repository files, issues, documentation or web pages may encounter prompt injection: text written to redirect its behavior, such as instructions hidden in a code comment or a web page asking it to send out credentials. Sandboxing and network restrictions limit the damage, because a blocked outbound request cannot leak data. They do not remove the risk entirely, since a manipulated agent can still make a harmful edit inside the workspace that a reviewer might overlook. Common mitigations include keeping secrets out of the agent's environment, restricting network access, reading diffs before merging and using read-only mode for unfamiliar repositories. Cloud tasks have their own separation. According to one third-party analysis, setup scripts may access the internet and secrets, while the agent phase gets neither by default.

Security scanning

OpenAI introduced Codex Security in March 2026 as a research preview after a private beta under the name Aardvark. It connected to GitHub repositories, built an editable threat model, attempted to reproduce suspected vulnerabilities in isolated environments and proposed patches for human review. At its DevDay event in September 2026, OpenAI announced Codex Security Cloud, which can scan whole repositories on demand, on a schedule or when new commits arrive. It investigates findings, removes duplicates and prepares fixes in the cloud, even when the user's laptop is closed. According to OpenAI's help documentation, fixes are proposals rather than automatic changes. The user reviews them and creates a draft pull request. Scans run in an ephemeral container that clones the repository, and the container is discarded afterward. Billing is token-based, existing customers must opt in before paid usage starts, and scanning pauses when funds are unavailable. A reviewer has noted that OpenAI's documentation does not specify the GitHub permissions the service requires. Third-party reviews that report detection counts or accuracy are not independent benchmarks, and their figures should be treated with caution.

Data handling

How code and conversations are treated depends on the account type. Local workflows run on the user's device, while cloud tasks run in OpenAI-managed environments. ChatGPT's training-data controls also cover content processed through Codex, including screenshots taken during computer use. On consumer plans, OpenAI's terms allow content to be used for training unless the user opts out, according to a third-party summary of ChatGPT plans. Plus and Pro accounts fall under consumer data terms and have no administrative features. Business workspaces, including Codex use within them, are excluded from training by default, and their data is encrypted in transit and at rest. Enterprise and Edu workspaces add administrative controls such as retention settings, audit logs, customer-managed keys and data residency options. Third-party guides also cite certifications such as SOC 2 and ISO 27001, and a HIPAA business associate agreement that covers only named models, so regulated users should verify coverage for each model. These training defaults are vendor statements and may change. Retention periods for connected services are set by those services, not by Codex.

Open-source ecosystem and development cadence

Licence scope and the open/closed split

The client side of Codex is published as open source. The CLI lives in a public repository under the Apache 2.0 licence and is written mostly in Rust. The repository's own README states the same licence. Apache 2.0 is a permissive licence, so the code can be read, modified and redistributed, including in commercial products.

The boundary of the open part is less uniform across sources. One encyclopedic infobox lists the CLI under Apache 2.0 and the desktop app as proprietary, while a parallel entry gives only the Apache 2.0 licence. OpenAI's developer material describes the open layer as the harness and integration surface, and says that model access and managed services remain separate. Third-party guides go further and state that the models, the IDE extension and Codex Cloud are not open source. In practice, the open code covers the client and the agent loop; the underlying models, hosted services and, according to the infobox, the desktop application are not part of it. The agent logic is organised in a component that is both a library where all the agent code lives and a runtime that runs one conversation thread, which is why the loop itself can be inspected.

Integration points for third parties

Several entry points let outside software build on Codex. The non-interactive mode can run a bounded agent workflow and return structured output for scripts, CI jobs and background tasks. The official SDK is aimed at application code that must start, resume or stream tasks, and OpenAI's documentation lists uses such as controlling Codex as part of a CI/CD pipeline and building Codex into internal tools. Official SDKs exist for TypeScript and Python, whereas Go has no first-party SDK, and community projects fill that gap.

For products that need deeper control, the app server is the relevant layer. It lets an application connect to a local Codex process, keep conversations open, stream events, interrupt work, expose tools, and respond to approval requests. OpenAI states that its own desktop app, terminal interface and web runtime, as well as third-party integrations such as JetBrains IDEs, VS Code and Xcode, all communicate with the harness through this layer. Codex can also be exposed as a tool for other agent frameworks: running the CLI in MCP-server mode exposes two Codex tools, which an orchestration layer can call.

OpenAI describes these components as open-source publications, the Codex CLI, app-server, and official Codex SDK. This is a vendor statement; the degree to which third parties actually depend on the app server in production is not established by the sources reviewed.

Release cadence

Releases are frequent. A third-party guide reports that OpenAI ships a new version every few days, with version 0.160.0 released on October 1, 2026. Subsequent patch releases followed within days, and the changelog entries list individual fixes with GitHub pull-request numbers. Release notes also change how users work. The April 2026 CLI release deprecated --full-auto and steered users toward explicit permission profiles and trust flows, which altered a commonly used approval setting.

Community contributions and feedback

The contribution model is more restrictive than the licence alone suggests. In the August 2026 release cycle, the project clarified that external contributions should go through issues and design discussion rather than pull requests. Community input therefore reaches the project mainly through bug reports, feature discussion and the public changelog, rather than through unsolicited code. Forks also exist; one widely cited fork describes itself as a fork of the original OpenAI Codex CLI with expanded model support and changed installation instructions.

Adoption figures are reported inconsistently. Star counts quoted by different aggregators range from roughly 119k to 127,800 stars and 20,000 forks as of October 2026, depending on when each page was captured. A secondary report from March 2026 gives 400 contributors and 640 tagged releases, but these numbers come from aggregator articles and have not been checked against the repository directly.

Debates and open questions

Several disputes about coding agents remain unresolved. They concern labor markets, how productivity should be measured, ownership of generated code, the governance of agents that act beyond a repository, and the gap between benchmark scores and results in practice. Much of the evidence is recent, and some of it comes from vendors or secondary reports.

Entry-level developers and employment

The most public debate asks whether coding agents are shrinking the pipeline of junior programmers. A Stanford study based on ADP payroll data used employment data provided by Automatic Data Processing Inc. and reviewed anonymized payroll information from millions of workers across tens of thousands of companies. It reported that workers aged 22 to 25 hold 20% fewer entry-level programming jobs than three years ago. The 2026 AI Index from Stanford HAI reportedly repeated the near-20 percent figure and noted that developers aged 30 and older in the same companies saw employment grow 6 to 12 percent over the same period.

Attribution is contested. The study's authors examined whether the job losses may have been caused by factors other than companies' adoption of AI. The lab's own later overview describes the aggregate effect on total employment as likely small right now, which points to a concentrated effect on specific entry-level roles rather than a broad contraction. Forecasts go further. Anthropic's chief executive has argued that AI could wipe out half of entry-level white-collar jobs over the next five years, but this is a projection and not a measured outcome.

Measuring productivity

Measurement is difficult, and the best-known controlled study points in an unexpected direction. METR's randomized trial in 2025 assigned issues from experienced open-source contributors to conditions with or without AI tools. Before starting tasks, developers forecast that allowing AI will reduce completion time by 24%. After completing the study, developers estimate that allowing AI reduced completion time by 20%. Measured completion times, however, rose: allowing AI actually increases completion time by 19%. The tools used were mainly Cursor Pro with Claude 3.5/3.7 Sonnet, which predate the Codex agent generation, so the result does not transfer directly.

METR's follow-up, begun in August 2025 with a larger pool of developers, is more tentative. For the returning subset, it now estimates a speedup of -18% with a confidence interval between -38% and +9%. METR also cautions that its developers and repositories don't represent most software development. Self-reported gains, including many figures cited in product marketing or in vendor announcements about Codex, should be read as perceptions unless independently measured.

Ownership, licensing and copyright

Two questions are often conflated. The first is whether AI-generated output can be owned at all. In the United States, copyright has been tied to human authorship, and a federal appeals court has upheld that requirement for a work generated autonomously by an AI system. Whether a developer who prompts an agent contributes enough human authorship to claim protection for the resulting code is a question that remains open and depends on the degree of human creative input.

The second question is whether training on publicly available repositories infringes rights or breaches open-source license terms. Litigation against GitHub Copilot, filed in 2022, argued that code was reproduced without the attribution and notice that many open-source licenses require. Outcomes differ across jurisdictions and are changing, so current rulings should be checked directly. A related issue is the licensing split: the Codex CLI is released under Apache 2.0, while the desktop and cloud components are proprietary. Which terms attach to a patch produced by a cloud task, when that patch may echo upstream code, has no settled answer in practice. See copyright, intellectual-property and open-source-software.

Governance of a general desktop agent

Broadening a coding agent into a general desktop agent changes its risk profile. A repository-bound agent mainly acts through files, a shell and version control, where changes can be reviewed as diffs. Connectors to email, calendars and document storage, and computer use on macOS, add actions with external effects: a message sent to the wrong recipient, a file deleted, or data disclosed after the agent follows instructions hidden in a web page or email. This form of cybersecurity risk, known as prompt injection, is harder to control because the agent must read untrusted content to do its work.

Governance therefore shifts from reviewing code to controlling accounts, operating-system permissions, and actions that cannot be undone. Sandbox and approval defaults designed for code may not suffice for personal or business data, and organizations must decide how workspace policies extend to such data. Questions of liability for an agent's actions, and of how data-privacy obligations apply to information the agent reads, are not resolved by the tool's design alone. The same question arises for any AI agent with broad access.

Benchmarks versus real engineering work

Vendor benchmark results have proven fragile. In February 2026 OpenAI stopped reporting scores on SWE-bench Verified. Its audit of 138 hard problems found that 59.4% contain flawed test cases — either too narrow (enforcing specific implementation details) or too wide (testing functionality not described in the problem). The same review found that all tested frontier models (GPT-5.2, Claude Opus 4.5, Gemini 3 Flash) can reproduce exact gold patches and verbatim problem details, indicating training data exposure.

The replacement benchmark has also been challenged. Reports state that OpenAI retracted its subsequent recommendation to use SWE-Bench Pro after an audit estimated that roughly 30 percent of its tasks were broken. A separate audit reported that Pro's graders mis-graded roughly a third of trials. These are secondary reports, and some of the figures circulating have not been confirmed by Scale or OpenAI. Scores are also sensitive to the harness: agent scaffolding -- not model capability -- increasingly determines the score.

The gap between benchmarks and field outcomes is the central tension. METR summarized its own finding as AI slows down experienced open-source developers in our RCT, but demonstrates impressive benchmark scores and anecdotally is widely useful. Vendor claims about Codex models should be weighed against this evidence, and against measured results on the reader's own codebase.

References

1. OpenAI Codex (language model)en.wikipedia.org
2. OpenAI GPT Model Release Timeline - Model Lineage, ChatGPT and Codex Milestones, and Platform Availabilityhidekazu-konishi.com
3. OpenAI Codex CLIaiwiki.ai
4. OpenAI Codex: From 2021 Code Model to a 2025 Autonomous Coding Agentmedium.com
5. OpenAI Codex (AI agent)en.wikipedia.org
6. OpenAI Codex App Review (2026): Mac, Windows, vs CLIvibecoding.app
7. How to make the Codex app work on Windows/Linux - CometAPIcometapi.com
8. OpenAI launches a Codex desktop app for macOS to run multiple AI coding agents in parallelventurebeat.com
9. Codex CLI - Review & Setup Guide - Developers Digestdevelopersdigest.tech
10. How OpenAI Codex Works, the codex-1 Coding Agenttheaiengineer.substack.com
11. GitHub Copiloten.wikipedia.org
12. GitHub, OpenAI release GPT-3-autocomplete combo for programmerstechbrew.com
13. OpenAI Codex Deprecation: Date, Reason, and Replacementbenchlm.ai
14. Codex Timeline: Deprecation (2023) to Agent Returnkunalganglani.com
15. Under the hood: Exploring the AI models powering GitHub Copilot - The GitHub Bloggithub.blog
16. Codexlablab.ai
17. GitHub Copilot And The Unfulfilled Promises Of An Artificial Intelligence Futurehackaday.com
18. What Is OpenAI Codex? The Complete History & Guide (2026)taskade.com
19. OpenAI Codex: Will It Replace Programmers?voiceflow.com
20. The Return of Codex AI — as an Agent -- Visual Studio Magazinevisualstudiomagazine.com
21. OpenAI Codex CLI — Open-Source AI Coding Agent (2026)llmreference.com
22. Introducing Codexopenai.com
23. Codex Version 3 Rumors and the Cloud-Agent Shiftvolanea.com
24. Codex (AI agent)en.wikipedia.org
25. OpenAI’s Codex Graduates to General Availability With New Slack Integration and Developer SDKwinbuzzer.com
26. Codex is now generally availableopenai.com
27. OpenAI Developers on X: "Codex is now GA, along with 3 features that make it more useful for engineering teams: - @Codex in Slack - Codex SDK - New admin tools" / Xx.com
28. Everything About Codex (2026): The Complete Guide to OpenAI’s Coding Agentbhavishyapandit9.substack.com
29. OpenAI Codex: Transforming Software Development with AI Agentsdevops.com
30. Codex Cloud Environments: Setup Scripts, Container Caching, Secrets, and the codex-universal Imagecodex.danielvaughan.com
31. How to Use Codex Cloud: A Guide to Setup, Environment Configuration, and Pricing Plans｜ソウタ｜経営・マーケに使える最新AI解説note.com
32. OpenAI Codex Tutorial: A Guide With 3 Practical Examplesdatacamp.com
33. Codex Cloud (2026): How It Works, Limits, Always-On Optionsagent37.com
34. Codex CLI: Setup, Sandbox Modes, and Verifying Its Outputtestmuai.com
35. Codex CLI Rust Migration Playbook: Config Changes 2026digitalapplied.com
36. OpenAI Codex Ships 90+ Plugins with MCP Servers Inside - Zuplozuplo.com
37. Codex CLI Plugin System: Bundling Skills, MCP Servers, and App Connectorscodex.danielvaughan.com
38. OpenAI Codex and Private Repositories: What Security Teams Should Knowaurascape.ai
39. OpenAI Codex CLI Cheat Sheet: Shortcuts and Commandscomputingforgeeks.com
40. Codex CLI sandbox modes: read-only, workspace-write, and when each is rightbackgrind.com
41. Codex CLI Deep Dive: Config, Profiles, Sandbox 2026digitalapplied.com
42. 6 best Codex plugins to use in 2026 - Vercelvercel.com
43. Model Release Noteshelp.openai.com
44. Introducing GPT-5.2-Codexopenai.com
45. Introducing GPT-5.2-Codex - Codex - OpenAI Developer Communitycommunity.openai.com
46. OpenAI: GPT-5.2-Codex Reviewdesignforonline.com
47. GPT-5.3 Codex: From Coding Assistant to General Work Agentdatacamp.com
48. GPT-5.3-Codexen.wikipedia.org
49. OpenAI Releases GPT-5.3-Codex, a New Codex Model for Agent-Style Developmentlaravel-news.com
50. ChatGPT & Codex changelogdevelopers.openai.com
51. OpenAI Launches GPT-5.2-Codex for Secure Codingesecurityplanet.com
52. GPT-5.2-Codex - API Pricing & Benchmarksopenrouter.ai
53. GPT 5.2 Codex Benchmarks, Cost and Capabilitiesvals.ai
54. GPT-5.3-Codex is now the base model for Copilot Business and Enterprise - GitHub Changeloggithub.blog
55. Release Notesopenai.com
56. GPT-5.2en.wikipedia.org
57. GPT-5.1en.wikipedia.org
58. OpenAI Release Notes - October 2026 Latest Updates - Releasebotreleasebot.io
59. OpenAI DevDay 2026 Recap for Developers - InfoQinfoq.com
60. OpenAI has significantly revamped the 'Codex CLI' and introduced 'Codex Cloud,' which allows Codex to run on the cloud. - GIGAZINEgigazine.net
61. OpenAI expands Codex and its API at DevDay with security scans, a Decisions API, and Ultrafastthe-decoder.com
62. OpenAI Gives Codex a Major Upgrade With Cloud Environments, Voice CLI & Code Reviewwindowsreport.com
63. OpenAI DevDay 2026 Recap for Developersdaily.dev
64. Codex Weekly: Cloud, CLI Voice, and Agents APIbighatgroup.com
65. OpenAI DevDay 2026: Dots, GPT-6.1 Sol & Codex Cloud Recapaiidelist.com
66. Introducing the Codex appopenai.com
67. New Codex App... Have you tried it yet? - Codex - OpenAI Developer Communitycommunity.openai.com
68. Codex's Code Review sends a comment to GitHub before you submit the reviewmixed-news.com
69. Review GitHub pull requests with Codexlearn.chatgpt.com
70. Codex Appcodex-docs.com
71. Codex Review (Verified August 2026)omidsaffari.com
72. Codex Pricingchatgpt.com
73. Using Codex with your ChatGPT planhelp.openai.com
74. Codex Pricing 2026: Free, $8 Go, $20 Plus, $100-$500 Pro, Business Plans, Credits, Usage Limitsmorphllm.com
75. OpenAI Codex pricing in 2026: every plan, cost, and gotcha explainedeesel.ai
76. Codex Pricing 2026: Plans, Credits, and DevDay Changesatoms.dev
77. Codex Usage Limits Explained: 5-Hour Reset, Weekly Quota, and Creditsknightli.com
78. GPT-5-Codex Modeldevelopers.openai.com
79. GPT 5.3 Codex API Pricing 2026 - Costs, Performance & Providerspricepertoken.com
80. Codex CLI Subscription Options: Every Plan, Limit and Cost (2026)inventivehq.com
81. Codex pricing: what you actually pay in 2026aialleyway.com
82. OpenAI Codex pricing in 2026: plans, token costs, and usage limitscloudzero.com
83. Codex Pricing in 2026: Free, Plus, Pro, API Costsomidsaffari.com
84. Codex Pricing (2026): Free, $8 Go, $20 Plus, Pro from $100layer3labs.io
85. OpenAI Codex Pricing 2026: Plans, Credits, Rate Card, and Usage Limits Explaineduibakery.io
86. OpenAI Codex Review: Is It Worth It in 2025?zackproser.com
87. OpenAI Codex App: A Guide to Multi-Agent AI Codingintuitionlabs.ai
88. Use subagents and custom agents in Codexsimonwillison.net
89. Subagentsdevelopers.openai.com
90. Early Takes on GPT-5-Codex: Developer Insights and Challengesblog.promptlayer.com
91. Codex in the Real World: Where It Works, Where It Breaks, and What It Means for Developersblog.whiteprompt.com
92. Developers Share Feedback on OpenAI's Codex Desktop App / Xx.com
93. Claude Code Competitors in 2026: A Complete Compare and Contrast - Claude Directory Blogclaudedirectory.org
94. Best AI Coding Assistants 2026: Cursor vs Copilot vs Claude Codescrimba.com
95. [AINews] OpenAI Codex App: death of the VSCode fork, multitasking worktrees, Skills Automationslatent.space
96. Agent approvals & securitydevelopers.openai.com
97. Sandbox and Approval Policiesdeepwiki.com
98. Codex Permissions: Sandbox, Approvals, Profilesturboai.dev
99. Securing OpenAI Codex: Sandbox Modes, Approval Policies, and the Two-Phase Runtime (2026)anomity.ai
100. Codex sandboxing: separate filesystem scope from approvals — harnezharnez.ai
101. Codex Sandbox Modes: Set Up Auto Mode in 3 Waysinventivehq.com
102. Codex Securityhelp.openai.com
103. Cloud FAQdevelopers.openai.com
104. OpenAI gives Codex reusable cloud environments that work across devicestechcrunch.com
105. OpenAI open-sources Codex Security CLI for repository scans and CIruntimewire.com
106. Is OpenAI's Codex Security Cloud safe to scan your repos?greenlitbooks.com
107. GitHub - openai/codex-security: OpenAI's Codex Security CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities. npm: https://www.npmjs.com/package/@openai/codex-security · GitHubgithub.com
108. Codex Security Review 2026 — OpenAI's AI That Fixes Your Vulnerabilitiesagent-finder.co
109. Managing data, sharing, and privacy in ChatGPT Businesshelp.openai.com
110. ChatGPT Work cloud securitydevelopers.openai.com
111. OpenAI Codex for Business: Teams & Enterprise Guidelayer3labs.io
112. ChatGPT and Codex: plans, models, admin and EU datalindamohamed.com
113. OpenAI Codex CLI (2026): Is It Free? Open Source, Latest Version & Installvibecoding.gallery
114. Codex as a platform: build on the open agent harnessdevelopers.openai.com
115. Unlocking the Codex harness: how we built the App Serveropenai.com
116. Codex SDKlearn.chatgpt.com
117. The Codex Go SDK Ecosystem: Embedding Coding Agents in Go Applicationscodex.danielvaughan.com
118. Building Consistent Workflows with Codex CLI & Agents SDKdevelopers.openai.com
119. Codex Changelog April 2026: Goals, Browser Use, GPT-5.5, and Safer Agentsdevelopersdigest.tech
120. Codex CLI Changelog (October 2026)gradually.ai
121. GitHub - ymichael/open-codex: Lightweight coding agent that runs in your terminal · GitHubgithub.com
122. Best Open Source CLI Coding Agents in 2026pinggy.io
123. OpenAI Codex CLI: The Complete Guidegetorchestra.io
124. OpenAI Codex CLI ships v0.116.0 with enterprise features: what developers should knowaugmentcode.com
125. GitHub - openai/codex: Lightweight coding agent that runs in your terminal · GitHubgithub.com
126. codex/README.md at main · openai/codex · GitHubgithub.com
127. ChatGPT & Codex changeloglearn.chatgpt.com
128. Using third-party agents in VS Codecode.visualstudio.com
129. Codex CLI Guide 2026: Setup, Sandbox, AGENTS.md & MCPblakecrosley.com
130. Stanford study finds AI has reduced availability of entry-level programming jobs - SiliconANGLEsiliconangle.com
131. AI jobs: devs under 26 lost 20% of work since 2022crypto.news
132. AI and Labor Markets: What We Know and Don't Know - Stanford Digital Economy Labdigitaleconomy.stanford.edu
133. It’s official. AI is already killing entry-level jobssfstandard.com
134. Measuring the Impact of Early-2025 AI on Experienced ...metr.org
135. AI coding tools make developers slower but they think they're faster, study findstheregister.com
136. We are Changing our Developer Productivity Experiment Design - METRmetr.org
137. METR 2026: AI Made Developers 19% Slower. Is That Still True?ingenire.com
138. Why SWE-bench Verified no longer measures frontier coding capabilitiesapp.daily.dev
139. Engineering Reliable Coding Agents: Evaluating and Operating the System Around the Modelarxiv.org
140. OpenAI drops SWE-Bench Verified, backs Pro after flaw auditaiweekly.co
141. SWE-bench Pro Leaderboard (September 2026): Every Model Score, Benchmarks, and Price per Pointmorphllm.com
142. Is SWE-bench Verified Contaminated? OpenAI Shifts to SWE-bench Procodesota.com
143. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity - METRmetr.org
144. Building Coding Benchmarks that Resist Contaminationimerit.ai