OpenAI introduced a new model, GPT-5.3 Codex, on February 5, 2026. This model is a major step forward in AI technology
What is GPT-5.3 Codex?
GPT-5.3 Codex is the most advanced agentic coding model OpenAI has created so far. Previously, we discussed the OpenAI Codex app, and now this new model powers the platform.
What does “agent” mean? This model doesn’t just write code; it can also handle long-running tasks, conduct research, use tools, and perform complex operations as a colleague would.
Agentic AI refers to artificial intelligence systems that make decisions and take actions independently to achieve specific goals. Unlike traditional AI, it can plan and act independently rather than waiting for constant user instructions.
Why is it so important?
Below are the key features that make GPT-5.3 Codex stand out:
- The First model that helped create itself.
This is an important development. GPT-5.3 Codex is the first model to play an active role in its own creation. The OpenAI Codex team used early versions of the model to:
- Debug its own training.
- Manage its own deployment process.
- Analyze test results and evaluations.
The model helped speed up its own development, representing an important milestone in artificial intelligence.
- Benchmark results
GPT-5.3 Codex achieved new records on industry benchmarks:
| Benchmark | GPT-5.3 codex | GPT-5.2 codex. | GPT-5.2 |
| SWE Benchmark Pro (Public) | 56.8% | 56.4 | 55.6% |
| Terminal Bench 2.0 | 77.3% | 64.0% | 62.2% |
| OS World verified | 64.7% | 38.2% | 37.9% |
| GDPval (wins or ties) | 70.9% | – | 70.9% |
Notice the OS World verified result, which improved from 38.2% to 64.7%. This shows significant progress in the model’s ability to use computers in visual desktop environments. Humans score about 72% on this test, so the model is now approaching human-level performance.
3. 25% Faster
GPT-5.3 Codex now runs 25% faster than earlier models because of upgrades to its infrastructure and inference stack, which means you get quicker exchanges and results.
Cybersecurity Capabilities
GPT-5.3 Codex is the first model to earn a high cybersecurity rating under OpenAI’s Preparedness Framework. This shows it is very effective in finding security vulnerabilities.
Cyber Range Performance
In OpenAI’s Cyber Range evaluation, GPT-5.3 Codex reached an 80% success rate. This is a big improvement over the previous model, GPT-5.1 Codex Max, which had a 60% success rate.
The model succeeded in the following scenarios:
- Azure SSRF attacks.
- Binary exploitation.
- Firewall evasion.
- Privilege escalation.
- Command and Control (C2) operations.
Trusted Access for Cyber (TAC) Program
OpenAI launched the Trusted Access for Cyber program to support defensive security researchers. The program supports use cases such as:
- Penetration testing.
- Red Teaming.
- Vulnerability assessment.
- Malware Reverse Engineering.
- Cryptographic Research
Web Development Capabilities
GPT-5.3 Codex can do more than write code; it can also create complete games and applications. OpenAI asked the model to develop two games to show what it can do:
- Racing Game is a comprehensive game with different racers, 8 maps, and items usable with the space bar.
- Driving Game: A game where you explore assorted reefs, collect fish, and manage oxygen and pressure.
The model created these games on its own, improving them step by step over the course of millions of tokens.
Interactive Collaboration
With GPT-5.3 Codex, you can interact with the model in real time as it works. You can ask questions, discuss ideas, and guide it towards solutions without losing context.
While the model is working:
- It provides frequent updates.
- Shares, key decisions, and progress.
- Response to feedback.
- Keeps you informed from start to finish.
Security and Safeguards
OpenAI has considered the potential risks of this powerful model. Here are some steps they have taken to address them:
Model Safety Training
- Ability to handle dual-use requests
- Refusal or de-escalation for harmful actions
- Restrictions on topics like malware creation and credential theft
Sandbox environment
- Network access is restricted by default.
- File edits are limited to the current workspace only.
- Native sandbox support for Windows, Mac OS, and Linux
Motoring and oversight
- Two-tier monitoring system
- Detection of high-risk usage
- Account-level enforcement
NVIDIA Partnership
GPT-5.3 Codex was built and runs on NVIDIA GB200 NVL72 systems. Working with NVIDIA improves the model’s performance.
Where Can You Access It?
You can use GPT-5.3 Codex now if you have a paid ChatGPT Plan:
- Codex App
- Codex CLI
- IDE Extension
- Web interface
OpenAI is still working to make the API access safe. The API will be available soon.
Conclusion
GPT-3.5 Codex is a big step forward for AI-powered coding. It can improve itself, is strong in cybersecurity, interactive, and 25% faster than before.
OpenAI says Codex is moving beyond writing code to do nearly anything developers and professionals can do on a computer. This model could be a real breakthrough for people in software development, design, product management, and data science.
What do you think? Would you like to try GPT-5.3 codecs? Share your thoughts in the comments below.
How To Try GPT-5.3 Codex? Codex App Waitlist
Do you want to try GPT-5.3 Codex? You will need a paid ChatGPT plan, such as Plus, Pro, Business, Enterprise, or Edu. You can also join the OpenAI Codex app waitlist to get early access.
Source: What is GPT-5.3-Codex? OpenAI’s Most Powerful Coding Agent
