OpenAI GPT-5-Codex Release and Codex Platform Updates

OpenAI has introduced GPT-5-Codex, a specialized version of GPT-5 optimized for agentic software engineering. This update transforms Codex from a coding assistant into a teammate capable of independently executing long-term tasks, conducting deep code reviews, and integrating across local and cloud environments.

GPT-5-Codex: Agentic Software Engineering

GPT-5-Codex is specifically trained for real-world software engineering workflows, including building full projects from scratch, debugging, performing large-scale refactors, and adding features and tests. Unlike general-purpose models, it is designed to be more steerable and adheres more strictly to AGENTS.md instructions.

Dynamic Reasoning and Performance

GPT-5-Codex dynamically adjusts its reasoning time based on task complexity. This allows the model to remain responsive during interactive chat sessions while dedicating more time to complex tasks.

Long-term Autonomy: In testing, the model has worked independently for over 7 hours on complex tasks, iterating on implementations and fixing test failures.
Token Efficiency: For the bottom 10% of user turns (sorted by model-generated tokens), GPT-5-Codex uses 93.7% fewer tokens than GPT-5. For the top 10% of complex turns, it spends twice as long reasoning, editing, and testing.
Benchmarks: OpenAI now reports results on all 500 SWE-bench Verified tasks, having resolved previous infrastructure limitations that limited reporting to 477 tasks.
Refactoring: The model is evaluated on refactor-style tasks from large repositories in Python, Go, and OCaml, including a Gitea pull request involving 232 files and 3,541 lines of code.

Code Review Capabilities

GPT-5-Codex is trained to identify critical flaws by navigating codebases, reasoning through dependencies, and executing code and tests to validate correctness. It matches the stated intent of a pull request (PR) to the actual diff, moving beyond the capabilities of static analysis tools. At OpenAI, the model now reviews the majority of internal PRs, catching hundreds of issues daily.

Front-end and Visual Integration

The model shows significant improvements in human preference evaluations for mobile websites and desktop apps. When operating in the cloud, GPT-5-Codex can process images or screenshots as input, visually inspect its own progress, and attach screenshots of the final result to tasks or GitHub PRs.

Codex Platform Updates

OpenAI has unified Codex into a single product experience connected via ChatGPT accounts, allowing seamless movement between local and cloud environments.

Codex CLI and IDE Extension

Codex CLI: The open-source CLI has been rebuilt for agentic workflows. It now supports image attachments (screenshots, wireframes), a progress-tracking to-do list, web search, and Model Context Protocol (MCP) for external system connectivity. Terminal UI updates include better formatting for tool calls and diffs, and three simplified approval modes (read-only, auto, and full access).
IDE Extension: Now available for VS Code, Cursor, and other VS Code forks. The extension allows developers to preview local changes and move work between the cloud and local environments without losing context.

Codex Cloud

Cloud infrastructure performance has been improved by caching containers, reducing the median completion time for new tasks and follow-ups by 90%. The cloud agent can now automatically set up environments by scanning for setup scripts and executing them, with configurable internet access for fetching dependencies via commands like pip install.

Security and Safety

To prevent data exfiltration and misuse, Codex runs in a sandboxed environment with network access disabled by default. Security features include:

Permission-based Actions: The agent can request permission before performing dangerous actions.
Customizable Access: Developers can limit cloud network access to trusted domains or grant the CLI/IDE extension access to web search and MCP servers.
Domain Safeguards: GPT-5-Codex is classified as "High capability" in Biological and Chemical domains, with specific safeguards implemented to mitigate associated risks.

Pricing and Availability

GPT-5-Codex is available to ChatGPT Plus, Pro, Business, Edu, and Enterprise users. As of September 23, 2025, it is also available to developers using Codex via API key through the Responses API, priced identically to GPT-5.

Plus, Edu, and Business: Support a few focused coding sessions per week.
Pro: Supports a full workweek across multiple projects.
Business/Enterprise: Offer credit-based scaling for additional usage needs.