OpenAI Codex (2026) is an autonomous coding agent built directly into ChatGPT, fundamentally different from the original 2021 Codex code-completion API. When given a natural language task description, Codex clones a GitHub repository into a network-isolated sandbox environment, reads the entire codebase for context, writes and modifies code across multiple files, installs dependencies, runs the test suite, iterates on failures, and opens a pull request — all without human intervention during execution. The agent is powered by the GPT-5.6 model family, with Sol as the flagship tier for complex reasoning and Sol Ultra (rolling out mid-2026) adding cooperative subagents and Cerebras-hardware acceleration for faster inference. Codex is available exclusively through ChatGPT subscriptions: Pro ($200/month, generous Codex usage included), Team ($25/user/month, limited access), and Enterprise (custom). There is no standalone Codex API or per-task billing — usage is bundled into the ChatGPT subscription. Key capabilities include multi-agent worktrees for parallel task execution across cloud and local environments, AGENTS.md configuration files for project-specific conventions and test commands, background scheduled automations for routine maintenance, and Skills for reusable workflow patterns. The sandboxed execution model provides strong security: no internet access during execution, preventing supply chain attacks and code exfiltration. Benchmark performance is exceptional on terminal-shaped agentic work — GPT-5.6 Sol scores 88.8% on Terminal-Bench 2.1 and 91.9% in Ultra mode, holding the published record. On repository-level issue resolution (SWE-bench Pro), it scores 64.6%, slightly behind Claude Code's 69.2%. User feedback emphasizes Codex's strength for asynchronous batch tasks: dependency upgrades, test coverage gaps, codebase-wide migrations, and documentation updates that can be queued and reviewed later. The primary limitation is lack of real-time interactivity — once a task starts, you cannot redirect or course-correct mid-execution, making it less suitable for exploratory feature development or live debugging compared to terminal-based agents like Claude Code.