What Codex is
Codex is OpenAI's agentic software development platform. It spans a desktop app, CLI, IDE extension, and cloud execution environment, and is powered by a dedicated lineage of GPT models optimized for long-horizon coding tasks. The product launched in May 2025 and has since grown into OpenAI's primary surface for software engineering — both externally, with 4 million weekly active users by April 2026, and internally, where output token volumes grew 13x–56x across OpenAI's own departments since November 2025.
The name has a dual history: the original Codex was a research model introduced in a 2021 paper that also established the HumanEval benchmark and the pass@k metric for evaluating functional correctness of generated code. That foundational work shaped how the field measures code generation quality. The current product is a distinct, much broader system that inherits the name but represents a full platform play rather than a standalone model.
Model lineage
Codex runs a dedicated model series that evolves in parallel with OpenAI's general-purpose GPT releases:
- GPT-5-Codex (September 2025): GPT-5 variant with dynamic thinking-effort adjustment, scaling compute by task complexity.
- GPT-5.1-Codex-Max (November 2025): optimized for project-scale, long-running autonomous work.
- GPT-5.3-Codex (February 2026): 25% faster than its predecessor, with improved real-time collaboration and in-session steering.
- GPT-5.3-Codex-Spark (February 2026): a latency-optimized variant exceeding 1,000 tokens per second via a partnership with Cerebras inference hardware; text-only, 128k context, research preview.
- GPT-5.4 (March 2026): first general-purpose model deployed across all Codex surfaces, adding native computer use and a 1M-token context window in experimental form.
- GPT-5.5 (April 2026): became the recommended Codex model, adding browser use and an automatic approval review agent.
- GPT-5.6 Sol/Terra/Luna (July 2026): current backbone, with Sol as the frontier tier featuring max reasoning and ultra mode (multi-subagent spawning), achieving 91.9% on Terminal-Bench 2.1.
Platform capabilities
Codex has expanded well beyond code completion into a set of agentic capabilities:
Computer use. Codex can see, click, and type within macOS and Windows applications. Remote computer use allows a locked machine to continue running tasks while the user monitors from a mobile device or another computer.
Goal mode. Graduated from experimental to generally available in May 2026, goal mode enables autonomous outcome-directed execution — the agent pursues a stated objective across multiple steps without per-step approval.
Browser use. The Codex app can operate an in-app browser for tasks like running local development servers and reproducing visual bugs.
ChatGPT Sites. Launched in preview in June 2026, Sites lets users build and deploy full-stack JavaScript/TypeScript web apps hosted by OpenAI, with ChatGPT-based authentication and RBAC controls — extending Codex from a code-generation tool into a lightweight internal app platform.
Workspace agents. Powered by Codex, workspace agents in ChatGPT automate complex multi-step workflows in the cloud for teams and enterprises.
Secure MCP Tunnel. An enterprise feature allowing Codex to connect to private or on-premise MCP servers via a customer-hosted tunnel without public internet exposure — a meaningful barrier removed for organizations running AI agents against internal systems.
Infrastructure and acquisitions
Two acquisitions signal the direction of Codex's infrastructure ambitions:
- Astral (March 2026): the company behind Ruff (Python linter) and uv (Python package manager), acquired to accelerate Codex's Python developer tooling through vertical integration.
- Ona (June 2026): a provider of secure, persistent cloud environments, acquired to support long-running agents across enterprise workflows.
OpenAI also disclosed that compute resources previously allocated to Sora were redirected to a new coding/enterprise model codenamed Spud, underscoring the strategic priority of the coding platform.
Distribution footprint
Codex reaches enterprise customers through multiple channels beyond the direct OpenAI API:
- AWS: OpenAI frontier models and Codex are generally available on Amazon Web Services, enabling deployment within existing AWS environments and procurement workflows.
- Cloudflare Agent Cloud: GPT-5.4 and Codex are integrated into Cloudflare's agent runtime for enterprise agentic workloads.
- Dell partnership: hybrid and on-premise deployments for infrastructure-sensitive enterprises.
- Microsoft Build: role-specific Codex plugins announced alongside Microsoft's own AI tooling.
- Codex Labs: a formal enterprise program with system integrator partnerships (Accenture, PwC, Infosys) to embed Codex across the software development lifecycle.
Samsung Electronics deployed ChatGPT Enterprise and Codex to employees worldwide, representing one of OpenAI's largest enterprise rollouts to date.
Research context and competitive position
The RAH (Recursive Agent Harness) paper (June 2026) used Codex as a baseline, reporting 71.75% on the Oolong-Synthetic long-context benchmark with a GPT-5 backbone — a number the RAH pattern raised to 81.36%, while Claude Sonnet 4.5 reached 89.77% in the same harness. This illustrates both Codex's role as a reference point for agentic coding research and the competitive gap that remains on long-context autonomous tasks.
The TraceLab dataset (UW, June 2026), drawn from ~4,300 real Codex and Claude Code sessions, characterizes the serving profile of production coding agents: long autonomous loops, long contexts with short outputs, heavily-tailed tool call distributions, and imperfect prefix cache hit rates. These findings motivate serving-system improvements — lower-overhead tool calling, append-length-aware prefill, better KV-cache management — that will shape the next generation of inference infrastructure for platforms like Codex.
The SkillOpt paper (May 2026) evaluated its skill-optimization framework inside the Codex agentic loop, lifting GPT-5.5 no-skill accuracy by up to +24.8 points — evidence that the Codex execution environment is becoming a standard harness for agentic research.
Product consolidation and trajectory
In July 2026, OpenAI released a unified ChatGPT desktop app for macOS and Windows that consolidates Chat, Work, and Codex into a single client — with Codex gaining inline editing, pull-request review in a sidebar, faster computer use powered by GPT-5.6, and multi-repository project support. Simultaneously, reporting described Codex being repositioned as a ChatGPT "superapp," signaling a consolidation from a specialized developer tool into a central surface of OpenAI's consumer and enterprise product strategy.
The trajectory is toward a platform that owns the full software development lifecycle: from code generation and review, through testing and deployment (ChatGPT Sites), to long-running autonomous agents operating across enterprise infrastructure — with dedicated model lines, acquired tooling, and a multi-cloud distribution network to match.
