OpenAI Codex in 2026: Models, Access, Pricing & Real Capabilities

The original Codex API is gone. Here’s what “Codex” means now, which GPT‑5.x‑Codex models it uses, how you access them, and what it’s actually good for in real repos.

Quick verdict on OpenAI Codex now Best for: teams that want an agentic coding environment to run multi-file edits, repo-wide refactors and scripted workflows on top of OpenAI’s latest coding models. Avoid if: you only need inline completions in VS Code or a single editor – GitHub Copilot, Cursor or IDE-native assistants will usually be simpler. If you’re still choosing your stack, look at a broader best AI coding tools shortlist before committing to Codex as your main agent. Starting price: Codex itself does not have a separate per‑token tariff; usage is billed through a mix of included credits in eligible ChatGPT plans and standard OpenAI API model rates when Codex executes API calls on your behalf. Exact prices depend on your ChatGPT plan and the underlying models used, and should be checked on the current ChatGPT and OpenAI API pricing pages. Main strength: agent orchestration over projects: multiple agents, tools and long-running workflows coordinated through a desktop app, web UI, IDE integrations, and CLI. For most founders this sits alongside tools like Claude Code vs Codex and IDE-first editors such as Cursor. Main limitation: cost and limits scale with the underlying models Codex uses; there is no separate low‑tier “legacy Codex” price band published by OpenAI. Model availability and limits depend on your ChatGPT plan, any Codex‑specific credits, and whether you are using Codex via ChatGPT sign‑in or with separate API keys. Where “OpenAI Codex” started vs what it is now The original Codex: a code-focused language model OpenAI first used the Codex name for a dedicated code model in 2021. The research described Codex as a GPT model fine‑tuned on public GitHub code, designed to translate natural language into source code and handle multiple programming languages and frameworks. arXiv GitHub Copilot’s technical preview in mid‑2021 was the first high‑profile product powered by this Codex generation, delivering autocomplete‑style suggestions directly in code editors. GitHub In OpenAI’s API, these models surfaced as endpoints like code-davinci-002 . Deprecation of the original Codex API Those 2021–2022 Codex models have since been deprecated. OpenAI’s current model list for the public API no longer includes code-davinci-002 and similar Codex LMs in its set of current production models. OpenAI documentation Today, there is no separate “Codex” language model line in the API comparable to that first generation. Instead, the Codex brand has moved up a layer. Codex now: a product layer on top of newer models “Codex” currently refers primarily to an agentic coding product: a coding agent available through ChatGPT’s web app, a desktop app for Windows and macOS, a CLI, and IDE integrations, which orchestrates OpenAI’s latest coding‑optimised models rather than a single standalone model. Wikipedia Under the hood, OpenAI documents specialised Codex models in the GPT‑5 family – notably GPT‑5.3‑Codex in its system card – and a smaller model called codex-mini-latest , which is a fine‑tuned version of o4-mini optimised for the Codex CLI. OpenAI documentation OpenAI Earlier GPT‑5‑series Codex variants are referenced in OpenAI’s deployment safety materials, but not all are listed as public, generally available API models. In parallel, OpenAI has published an addendum system card describing a codex-1 model built on its o3 series, indicating that the Codex brand has iterated over more than one underlying model generation before the current GPT‑5‑series Codex family. OpenAI documentation Current Codex architecture: product, models, and access paths Codex as desktop app, web agent, IDE integrations, and CLI OpenAI positions Codex as an agentic coding environment rather than a simple editor autocomplete. The product surface includes access through ChatGPT’s web app, a desktop app for Windows and macOS, a CLI for scripting workflows and automation, and several IDE integrations. Wikipedia GitHub Codex is documented as a desktop command centre spanning app, web, CLI and IDE Agents can spawn sub‑agents, invoke tools (e.g. shell, tests, linters) and operate over project‑scale context, according to recent research describing Codex as evidence of the wider “shift to agentic AI”. arXiv For a practical view of how this compares to other agent-first stacks, see the broader AI development workflow guide . Core model family under Codex OpenAI documents a dedicated coding model line in the GPT‑5 family. Primary OpenAI sources describe GPT‑5.3‑Codex as the current flagship Codex model and codex-mini-latest as a smaller, fast reasoning model fine‑tuned from o4-mini for use in the Codex CLI. OpenAI documentation OpenAI Deployment‑safety materials and secondary sources also reference earlier GPT‑5‑series Codex variants and a GPT‑5.3‑Codex‑Spark variant, but these are not all documented as separate public pricing SKUs. GitHub Wikipedia GPT‑5.3‑Codex is described in its system card as OpenAI’s most capable agentic coding model at the time of its release in February 2026, intended for software development workflows and governed by the same preparedness and safety safeguards that OpenAI applies to other high‑capability GPT‑5‑series models. OpenAI documentation How Codex uses OpenAI’s model list and pricing There is no dedicated Codex billing SKU for per‑token usage. Instead: Codex calls OpenAI models via standard API endpoints , including Codex‑branded and non‑Codex models from the public list, when it executes API requests on a user’s behalf. OpenAI documentation Per‑token usage from those API calls is billed at the standard per‑token rates for the underlying models, while some Codex usage is covered by Codex‑specific credits or included ChatGPT plan usage according to OpenAI’s current terms. OpenAI documentation This means Codex is effectively a product layer and orchestrator on top of the general OpenAI model API and its pricing, combined with Codex‑specific and ChatGPT‑plan entitlements. If you’re trying to forecast your bill, it’s worth cross‑checking against how ChatGPT pricing works overall in 2026 via the detailed ChatGPT pricing guide . Sign‑in options and impact on models Codex supports two main access patterns: OpenAI documentation GitHub ChatGPT sign‑in – using a ChatGPT account (such as Plus, Pro, or business tiers where available) as the identity and billing context, with Codex usage drawing on included credits and plan‑specific limits. API key / org credentials – connecting Codex directly to an OpenAI API key tied to an organisation, so that Codex’s API calls are billed to that organisation’s API account. Codex also allows configuration of external model providers that expose OpenAI‑compatible endpoints, including via OAuth‑style flows, which community guides describe as a way to route Codex’s calls through other model hosts when permitted. Reddit Community reports describe changes over time in which models appear when signing into Codex with a ChatGPT account, versus connecting with an API key; users report that some ChatGPT‑only options have been removed from the Codex model picker, while API key users can access the broader set of API‑listed models. These differences are not exhaustively documented in OpenAI’s public product pages and may vary by plan and date. Reddit Real capabilities: what Codex can actually do in 2026 Agentic coding over multiple steps OpenAI and independent researchers describe Codex as exemplifying agentic AI – systems that decompose tasks, plan multi‑step workflows and call tools rather than just predicting the next token. arXiv In practical terms, this translates into workflows such as: Plan and edit : analyse a task (for example, “add passwordless login”), plan a sequence of edits, then apply changes across several files. Tool‑augmented loops : run tests, linters and build steps; inspect failures; propose and apply fixes. Multi‑agent roles : one agent drafts changes, another reviews or writes tests, with orchestration handled by the Codex runtime. Long‑running sessions and desktop workflows The Codex desktop app is documented as a way to manage long‑running code workflows and multiple agents. Wikipedia OpenAI spokespeople have framed Codex as a long‑term product, comparable in importance to ChatGPT, used to orchestrate GPT‑5‑series models for day‑to‑day development work. Axios This design is aimed at workflows that go beyond short chat sessions: Maintaining context over entire sprints or epics , with persistent sessions tied to repositories. Running scheduled or triggered jobs , such as nightly refactors or dependency checks. Coordinating CI/CD hooks – for example, reviewing PRs, proposing fixes, or running migration scripts. Repository‑scale operations Because Codex sits on top of large‑context models like GPT‑5.3‑Codex, it can operate at repository scale rather than per‑file. Documentation and system cards emphasise project‑level understanding and workflows. OpenAI documentation Typical use cases discussed in public material include: Refactors – systematic renames, extraction of shared modules, pattern replacements across many files. Framework or version migrations – assisting with upgrades (for example, framework major versions), flagging breaking changes, proposing patches. Test generation – adding test coverage for existing codebases by analysing functions and classes. Glue and integration code – wiring APIs, SDKs, or infrastructure code where patterns are common. Where Codex is strong vs weak today Based on OpenAI’s positioning, system cards and comparative research, the strengths and weaknesses of Codex in 2026 can be summarised as: Strong for : Greenfield feature work where the task can be expressed clearly in natural language, with the model handling scaffolding and boilerplate via GPT‑5.3‑Codex. Repo‑wide pattern changes , which benefit from project‑scale context windows. Glue code and integrations , where patterns are repetitive and well‑represented in training data. Migrations and test generation , especially when paired with automated test execution and feedback. Weaker for : Complex system architecture decisions and deep non‑code context (organisation‑specific constraints, legacy operational realities) that are hard to encode in prompts. Unreliable toolchains – flaky tests, non‑deterministic builds or unclear error messages can degrade agent performance, as noted in broader agentic AI analyses. arXiv Explaining and enforcing nuanced governance rules (for example, internal security policies) without additional guardrails or custom tooling. Models behind Codex: from GPT‑5.3‑Codex to the GPT‑5.6 family Current state first: Codex's own model documentation now recommends GPT‑5.6 Sol (flagship, strongest for complex coding), GPT‑5.6 Terra (balanced everyday work) and GPT‑5.6 Luna (fast and lowest cost), all available across the desktop app, web and Codex CLI. The same page lists gpt-5.2 and gpt-5.3-codex as already deprecated for Codex with ChatGPT sign-in, and schedules gpt-5.4 and gpt-5.4-mini to retire on 31 August 2026 , with gpt-5.6-terra and gpt-5.6-luna as their replacements. Codex authenticated with your own API key is not affected by that retirement. Codex docs Codex now recommends GPT-5.6 Sol, Terra and Luna, and carries a Deprecated Codex models section GPT‑5.3‑Codex in brief (background) The GPT‑5.3‑Codex system card describes it as a specialised large language model for agentic coding, combining frontier coding performance with tooling for software development workflows. OpenAI documentation Key points from the system card and related documentation include: Training focus on code, developer documentation and software artefacts. Agentic behaviour – designed to work with tools, sub‑agents and long‑context reasoning for coding tasks. Safety and policy constraints governed by OpenAI’s preparedness framework and safeguards for high‑capability GPT‑5‑series models, including restrictions on certain types of code generation. OpenAI documentation Other Codex‑aligned models OpenAI’s model references for Codex and related tooling mention several tiers of models: GitHub OpenAI documentation Earlier GPT‑5‑series Codex variants – including names such as GPT‑5‑Codex and GPT‑5.1‑Codex in deployment‑safety and research materials. These are referenced as part of the evolution of Codex but are not all exposed as distinct public API products with separate pricing. GPT‑5‑Codex‑Max and similar variants – higher‑capacity or higher‑limit variants referenced in internal and deployment‑safety documentation; public pricing details for these are limited and subject to change. GPT‑5.3‑Codex‑Spark – an ultra‑fast variant introduced as a research preview for real‑time interactive coding in the Codex app, Codex CLI, and IDE extensions, according to OpenAI’s announcement. Wikipedia codex-mini-latest – described by OpenAI as a “fast reasoning model optimised for the Codex CLI” , implemented as a fine‑tuned version of o4-mini with a set of snapshots and aliases. OpenAI documentation Exact latency and context window values are not publicly tabulated for each variant in the provided sources. However, model naming and placement alongside other small reasoning models suggest the usual trade‑off: larger GPT‑5.3‑Codex models for top‑end quality and smaller models such as codex-mini-latest for speed and cost. Choosing models inside Codex Within the Codex CLI and app, model choice works in layers: Default models are selected based on task type (for example, a default GPT‑5‑series Codex model for coding agents, and smaller models for lightweight reasoning). Per‑task overrides allow advanced users to specify models explicitly – for example, selecting a higher‑capacity GPT‑5.3‑Codex variant for a high‑stakes migration. Cost‑aware workflows can route sub‑tasks to cheaper models (such as codex-mini-latest or other small reasoning models) while reserving frontier models for final synthesis or more complex sections. The Codex repository and documentation show that non‑Codex models such as GPT‑4o‑mini‑class models can also be used inside Codex workflows for specific subtasks, reinforcing that Codex is a general orchestrator over the OpenAI model API. GitHub Pricing, limits, and how Codex really bills you No Codex‑specific per‑token tariff: everything flows from model prices and credits OpenAI’s help centre materials on using Codex with ChatGPT emphasise that Codex usage combines Codex‑specific and ChatGPT‑plan credits with standard API pricing for any additional underlying model usage, rather than having a fully separate Codex rate card. OpenAI documentation Codex has no separate seat: it is included in the ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise plans That means costs are determined by: The plan and credits associated with the user’s ChatGPT or Codex subscription. The model selected for any additional API calls (for example, GPT‑4o‑class models versus GPT‑5.3‑Codex). The tokens consumed (input + output) per API request beyond included credits. Reference pricing bands OpenAI’s public pricing is regularly cited in academic and technical documents when quantifying costs. For example, an ICLR 2026 paper cites per‑million‑token prices for GPT‑4o‑mini at US$0.15 input / US$0.60 output , and GPT‑4o at US$2.50 input / US$10.00 output as official rates at the time. OpenReview Wikipedia OpenAI’s model page for codex-mini-latest documents it as a small, fast reasoning model fine‑tuned from o4-mini for the Codex CLI. It is not listed as a separate SKU with its own public price line, but appears in the same tier as other small reasoning models; by analogy with GPT‑4o‑mini, this suggests pricing in a similar order of magnitude, although exact rates should be confirmed on OpenAI’s current pricing page. OpenAI documentation Plan / model As of Price Source GPT‑4o‑mini API (reference for low‑cost agents) 2025‑11‑01 US$0.15 / 1M input tokens, US$0.60 / 1M output tokens OpenReview GPT‑4o API (reference flagship pricing) 2024‑07‑18 US$2.50 / 1M input tokens, US$10.00 / 1M output tokens Wikipedia codex-mini-latest via API 2026‑08‑13 Not listed as a separate public SKU; billed at the same rate as other small reasoning models in its tier according to the current OpenAI pricing page OpenAI documentation Using Codex with ChatGPT Plus / Pro 2026‑04‑01 No explicit Codex‑only per‑token rate; usage draws on ChatGPT and Codex‑specific credits and is subject to ChatGPT limits, while any additional Codex‑initiated API calls bill at standard model rates OpenAI documentation Public documentation in the sources cited here does not provide a precise per‑token price for GPT‑5.3‑Codex itself. Frontier coding models are typically priced above smaller models like GPT‑4o‑mini, but any specific figure needs to be taken from the current OpenAI pricing page. Limits and caps Codex usage is constrained by: Model‑level limits : context window size, rate limits and per‑minute token caps defined at the API level for each model. OpenAI documentation Organisation‑level limits : Codex‑specific safety and usage caps that an organisation can configure, including project‑ and workspace‑level controls. ChatGPT plan limits if using Codex through ChatGPT credentials, which OpenAI directs users to cross‑check with the central pricing and usage documentation. OpenAI documentation Repository‑scale operations are possible given large contexts and agent loops, but they can become expensive in both tokens and credits. Open documentation and community discussion indicate a common pattern: reserving frontier models such as GPT‑5.3‑Codex for high‑value edits, while relying on cheaper models for background analysis or non‑critical tasks. OpenAI If you want a founder-level view of what these token bands mean for an actual product budget, see the breakdown in the real cost of an AI MVP . Practical cost patterns Exact costs depend heavily on project size and usage patterns. Using the reference bands above and OpenAI’s own guidance: Simple autocomplete / one‑file chat can often run on small models similar to GPT‑4o‑mini, costing well under a US dollar in API usage for many sessions, with some portion potentially covered by Codex‑specific credits. Single feature implementation using GPT‑5.3‑Codex across design discussions, code generation and tests can accumulate into the dollars‑to‑tens‑of‑dollars range in API usage, depending on iteration depth and how much is covered by included credits. Repo‑wide refactors or migrations may consume large token vol