OpenAI’s GPT-5 Codex is here

Estimated reading time: 3 minutes

Everything engineering teams need to know about OpenAI’s latest AI coding assistant enhancements

OpenAI has released GPT-5 Codex, with the latest version of the agentic coding assistant designed to take on more complex development tasks, from building projects end-to-end, to reviewing code changes. The updates look to bring its capabilities up to speed with GitHub’s May update for its own Copilot coding assistant.

Here’s what’s new, and what it means for engineering teams.

A model built for complex engineering tasks

The new version of Codex is a specialized version of the latest GPT-5 large language model, which was released in August to a mixed user response. It has been designed to:

• Build full projects from scratch.
• Add new features and tests.
• Debug existing code.
• Perform large-scale refactors.
• Review code changes.

According to OpenAI, the model is better at following instructions, and produces higher-quality code without developers needing to provide long prompts about style or cleanliness.

Smarter use of time and tokens

A key change is “dynamic thinking time”. Codex now adapts how long it spends working on something based on task complexity. For simple, well-defined requests or interactive sessions it aims to provide a quick response. For more complex tasks, such as large refactors or multi-step workflows it will extend runs, with the ability to work more independently.

In OpenAI’s own testing, the model has run independently for over seven hours at a time, iterating on implementations, fixing test failures, and delivering working results.

The company also highlights efficiency gains:

• On short interactions (bottom 10% of user turns), GPT-5-Codex uses 93.7% fewer tokens than GPT-5.
• On long, complex turns (top 10%), it spends about twice as long reasoning, editing, testing, and iterating.

Tooling upgrades: CLI, IDE, and cloud

OpenAI has rebuilt Codex’s developer tools depending on where users want to deploy these new capabilities.

Codex CLI

• Supports attaching and sharing images (e.g. wireframes, screenshots, diagrams).
• Tracks progress with to-do lists.
• Provides improved formatting of previews and diffs in the terminal.
• Introduces simplified approval modes with three levels of access.
• Stores conversation state more compactly, enabling longer sessions.
• Improved accuracy when using external tools like web search and the Model Context Protocol (MCP).

IDE extension

• Available for VS Code, Cursor, and VS Code forks.
• Lets Codex preview and edit code in context, reducing prompt length.
• Enables smooth transitions between cloud and local workflows: opening cloud tasks in IDEs, reviewing completed tasks, and creating new cloud jobs directly from the editor.

Codex cloud environment

• 90% faster task setup thanks to container caching.
• Automatic environment setup by scanning and running common setup scripts.
• Ability to fetch dependencies at runtime.
• Image inputs supported: developers can attach screenshots or design specs, and Codex can capture its own browser screenshots to show progress.

Better code reviews

Codex now comes with a dedicated code review mode that integrates with GitHub. It has been trained to:

• Catch critical flaws by navigating codebases, reasoning through dependencies, and executing code/tests to validate behavior.
• Provide review comments that are less likely to be incorrect or unimportant.
• Once enabled, Codex automatically reviews pull requests when they move from draft to ready.
• Developers can request Codex reviews explicitly with “@codex review” comments, optionally adding context such as “focus on security” or “check dependencies.”
• Codex can even implement its own suggested edits when asked.

Safety and trust

Codex has also been given some safety updates:

• By default, Codex runs in a sandboxed environment with network access disabled.
• Developers can configure security, such as allowing network access only to trusted domains.
• The system asks for confirmation before running potentially dangerous actions.
• Codex outputs diagnostics, including citations, terminal logs, and test results, to support human review.

OpenAI advises that developers review its output before deploying to production.

Pricing and availability

GPT-5-Codex is available today as part of:

• ChatGPT Plus, Pro, Business, Education, and Enterprise plans.
• Usage levels scale with subscription tiers, from light weekly sessions (Plus) to full-time workloads (Pro).
• Enterprise plans include pooled credits and the ability to extend beyond included limits.

The bottom line

With GPT-5-Codex, OpenAI is trying to moving Codex from a coding autocomplete tool toward a full-fledged development assistant. 

By combining dynamic reasoning, project-scale awareness, integrated tooling, and safer defaults, Codex aims to help developers spend less time on setup and repetitive fixes.

Get articles like this in your inbox

Choose your LeadDev newsletters to subscribe to.