Codex

Overview

GPT-5.3-Codex represents OpenAI's most significant expansion of Codex capabilities, released on February 5, 2026. This model transforms Codex from a specialized AI code generator into a general-purpose AI agent capable of handling the full spectrum of professional work—from software engineering to creating presentations, analyzing spreadsheets, and executing complex research tasks. Available to paid ChatGPT users across all Codex surfaces (app, CLI, IDE extensions, and web), GPT-5.3-Codex delivers 25% faster performance while consuming fewer tokens per task.

The model achieves state-of-the-art results across multiple benchmarks: 56.8% on SWE-Bench Pro (coding), 77.3% on Terminal-Bench 2.0 (terminal skills), and 64.7% on OSWorld-Verified (computer-use tasks). As covered in our AI agent tools comparison, this represents the highest computer-use capability among publicly available models. Notably, GPT-5.3-Codex was instrumental in creating itself—OpenAI's team used early versions to debug training, manage deployment, and diagnose evaluations, demonstrating its production-ready reliability.

What's New

Beyond Coding: General Computer-Use Capabilities

GPT-5.3-Codex marks the first Codex model designed to operate across the full spectrum of professional work, not just software engineering. The model matches GPT-5.2's 70.9% win rate on GDPval—an evaluation spanning 44 occupations including financial analysis, legal research, and business consulting. Users can now task Codex with creating PowerPoint presentations, analyzing complex spreadsheets, drafting reports, and conducting multi-step research workflows.

In OSWorld-Verified (a benchmark measuring visual desktop task completion), GPT-5.3-Codex achieves 64.7% accuracy compared to GPT-5.2-Codex's 38.2%—a 69% improvement. This enables Codex to navigate graphical interfaces, manage files, interact with AI productivity tools, and execute tasks that previously required human intervention.

Coding Performance and Efficiency Gains

GPT-5.3-Codex sets new benchmarks across software engineering evaluations while delivering substantial efficiency improvements:

• SWE-Bench Pro: 56.8% (up from 56.4% in GPT-5.2-Codex), maintaining SOTA performance on the industry's most rigorous multi-language evaluation spanning four programming languages
• Terminal-Bench 2.0: 77.3% (up from 64.0%), demonstrating superior terminal command execution and system interaction
• Cybersecurity CTF Challenges: 77.6% (up from 67.4%), marking the first model classified as "High capability" under OpenAI's Preparedness Framework
• SWE-Lancer IC Diamond: 81.4% (up from 76.0%), excelling at real-world contractor-level engineering tasks

Critically, GPT-5.3-Codex achieves these results with fewer tokens than prior models, reducing both cost and latency for users running batch operations or high-frequency automated workflows.

Web Development and Long-Horizon Building

Combining frontier coding with improved aesthetics and reasoning, GPT-5.3-Codex can build complex, production-ready web applications from scratch over multi-day sessions spanning millions of tokens. OpenAI demonstrated this capability by tasking the model with creating two full-featured games (a racing game with 8 maps and items, plus a diving game with reef exploration mechanics) using only generic follow-up prompts like "fix the bug" or "improve the game."

The model also delivers stronger default behaviors for underspecified prompts. When asked to build a landing page, GPT-5.3-Codex automatically implements smart UX decisions (like showing discounted monthly prices for yearly plans) and creates multi-testimonial carousels instead of single quotes, resulting in more polished first drafts.

Real-Time Steering and Interactive Collaboration

GPT-5.3-Codex introduces continuous interaction during task execution, fundamentally changing how users work with AI agents. Instead of waiting for final outputs, users can now:

• Ask clarifying questions mid-task without breaking context
• Discuss alternative approaches while Codex is working
• Provide real-time feedback to steer direction
• Receive frequent progress updates on key decisions

This shift transforms Codex from a "request → wait → receive" tool into an interactive colleague, enabling faster iteration cycles and reducing wasted work from misaligned assumptions. The steering mode is enabled by default in the stable release, with configurable interaction preferences available in the Codex app settings.

Cybersecurity: First "High Capability" Model

GPT-5.3-Codex is the first model OpenAI has classified as "High capability" for cybersecurity tasks under its Preparedness Framework. The model was directly trained to identify software vulnerabilities, achieving 77.6% on capture-the-flag challenges—a 10-percentage-point leap over GPT-5.2-Codex.

To support defensive use while mitigating misuse risk, OpenAI is deploying:

• Trusted Access for Cyber: Pilot program accelerating cyber defense research for vetted organizations
• $10M Cybersecurity Grant Program: API credits for open source and critical infrastructure security
• Enhanced monitoring: Automated detection pipelines with threat intelligence integration
• Comprehensive safety stack: Safety training, automated monitoring, and enforcement pipelines including threat intelligence

The model has been tested for vulnerability detection capabilities across real-world codebases, supporting defensive security research and open source security initiatives.

Infrastructure and Performance Optimizations

GPT-5.3-Codex runs 25% faster than GPT-5.2-Codex due to improvements in OpenAI's inference stack and co-design with NVIDIA GB200 NVL72 systems. This speed increase applies across all interaction points—code generation, debugging sessions, and multi-file refactors—translating to faster turnaround times for production deployments and reduced wait times during interactive sessions.

Availability & Access

GPT-5.3-Codex is available to users with paid ChatGPT plans through multiple access points:

• Codex Desktop App (currently macOS with Apple Silicon; Windows version coming soon)
• Codex CLI (command-line interface, cross-platform)
• IDE Extensions (VS Code, JetBrains, and other editors with platform-specific support)
• ChatGPT Web Interface (all platforms)

Additionally, limited-time trial access is available to ChatGPT Free and ChatGPT Go users—check the official Codex page for current availability.

API access is coming soon. OpenAI is working to enable programmatic access for developers building custom integrations, automated workflows, and enterprise deployment scenarios.

Cybersecurity Capabilities & Safety Framework

As the first model classified as "High capability" for cybersecurity under OpenAI's Preparedness Framework, GPT-5.3-Codex is deployed with comprehensive safeguards:

• Cautious deployment approach: Enhanced safety training, automated monitoring, and enforcement pipelines with threat intelligence integration applied to all users
• Trusted Access for Cyber program: Vetted security researchers and organizations can apply for specialized access to support defensive security research (application required)
• Grant program support: $10M in API credits available for open source maintainers and critical infrastructure defenders conducting good-faith security research

Pricing & Plans

GPT-5.3-Codex is included with paid ChatGPT subscriptions. No separate pricing tier exists for Codex-specific features.

What's Included with Codex Access:

• GPT-5.3-Codex for coding and computer-use tasks
• All Codex surfaces (app, CLI, IDE, web)
• Real-time steering and interactive collaboration
• Multi-hour session persistence
• Sandbox execution environments

API Pricing (coming soon): Token-based pricing details to be announced. Contact OpenAI sales for enterprise API access and volume discounts.

Pros & Cons

Pros

• True general-purpose agent — First Codex model capable of professional work beyond coding, handling PPTs, spreadsheets, research, and system administration with production-ready quality
• Performance across all benchmarks — Sets SOTA on SWE-Bench Pro (56.8%), Terminal-Bench (77.3%), and OSWorld (64.7%), proving reliability across diverse task types
• Interactive collaboration mode — Real-time steering eliminates the "black box" problem, letting you guide work without context loss or interruptions
• 25% faster with lower token consumption — Reduced latency and cost per task make high-frequency automation more economical
• Self-improving development — Proven capability: OpenAI used GPT-5.3-Codex to accelerate its own training and deployment, validating production reliability
• Cybersecurity leadership — First "High capability" model for vulnerability detection, backed by $10M grant program and Trusted Access initiative

Cons

• API access delayed — Programmatic access not yet available, limiting automation workflows and enterprise integrations requiring custom deployments
• Limited free access — Free and ChatGPT Go trial access is time-limited; sustained use requires Plus ($20/month) or higher paid plan
• High capability classification safeguards — Enhanced cybersecurity monitoring and safety measures may affect user experience for advanced security research workflows
• Desktop app platform limitations — Native Codex app currently macOS-only; Windows users must rely on CLI, IDE extensions, or web interface until desktop app expands
• Computer-use still evolving — OSWorld 64.7% performance shows room for improvement; complex GUI tasks may require multiple attempts or clarifications

Best For

• Software engineering teams managing complex, multi-repository codebases requiring cross-file refactors and long-horizon debugging sessions spanning hours or days
• Security researchers conducting vulnerability assessments, penetration testing, and defensive security automation with the most capable cybersecurity AI available
• Product managers and non-technical founders who need to prototype web applications, analyze user data, and create presentations without hiring developers or designers
• Data scientists and analysts working with complex spreadsheets, multi-step data pipelines, and automated report generation across multiple tools and formats
• DevOps engineers automating deployment workflows, infrastructure management, and terminal-based system administration tasks requiring adaptive problem-solving
• Technical writers and documentation teams generating code examples, API references, and technical diagrams with verified accuracy and executable samples

FAQ