GPT-5.3-Codex release notes / what's new
OpenAI has introduced GPT-5.3-Codex, a high-capability agentic coding model designed to handle long-running tasks involving research, tool use, and complex execution. The model combines the frontier coding performance of GPT-5.2-Codex with the reasoning and professional knowledge of GPT-5.2, while operating 25% faster than its predecessors.
Frontier Agentic Capabilities
GPT-5.3-Codex establishes new industry benchmarks in coding and computer-use tasks, moving beyond simple code generation to autonomous execution.
Software Engineering and Terminal Skills
GPT-5.3-Codex achieves state-of-the-art performance on SWE-Bench Pro, a multi-language evaluation of real-world software engineering, and significantly exceeds previous performance on Terminal-Bench 2.0. Notably, the model achieves these results using fewer tokens than prior models, increasing efficiency for users.
Web Development and Iterative Design
The model demonstrates advanced capabilities in building complex, functional games and applications from scratch. In internal tests, GPT-5.3-Codex autonomously iterated on a racing game and a diving game over millions of tokens using generic follow-up prompts such as "fix the bug" or "improve the game."
Additionally, the model shows improved intent understanding for day-to-day website creation. For example, when prompted to build a landing page for "Quiet KPI," GPT-5.3-Codex automatically implemented production-ready features—such as a discounted monthly price for yearly plans and a transitioning testimonial carousel—that were absent in GPT-5.2-Codex.
General Professional Knowledge Work
GPT-5.3-Codex supports the full software lifecycle, including debugging, deploying, monitoring, writing PRDs, and user research. Its capabilities extend to general knowledge work, matching GPT-5.2 on the GDPval benchmark, which measures performance across 44 different occupations, including the creation of spreadsheets and presentations.
In terms of computer-use capabilities, GPT-5.3-Codex shows a significant leap in performance on the OSWorld-Verified benchmark, where agents use vision to complete productivity tasks in a visual desktop environment.
Self-Improving Development Cycle
GPT-5.3-Codex is the first OpenAI model instrumental in its own creation. The Codex team utilized early versions of the model to:
- Debug its own training runs and track patterns during training.
- Manage deployment and diagnose test results.
- Optimize the harness for GPT-5.3-Codex and identify context rendering bugs.
- Dynamically scale GPU clusters to adjust to traffic surges and maintain latency stability.
- Analyze alpha testing logs using regex classifiers to estimate productivity and user response frequency.
Interactive Collaboration and Steering
To bridge the gap between model capability and human supervision, GPT-5.3-Codex is more interactive within the Codex app. Users can now steer the model in real time—asking questions and discussing approaches—while it works, rather than waiting for a final output. This "steering" behavior can be enabled in the app settings under General > Follow-up behavior.
Cybersecurity and Safety Framework
GPT-5.3-Codex is the first model classified as "High capability" for cybersecurity tasks under OpenAI's Preparedness Framework and the first directly trained to identify software vulnerabilities.
Safety Mitigations
Due to the dual-use nature of cybersecurity capabilities, OpenAI has implemented a comprehensive safety stack, including:
- Safety training and automated monitoring.
- Trusted access for advanced capabilities.
- Enforcement pipelines utilizing threat intelligence.
- Automatic routing of high-risk cyber requests from GPT-5.3-Codex to GPT-5.2 to prevent misuse.
Ecosystem Support
OpenAI is launching "Trusted Access for Cyber," a pilot program for defense research. They are also expanding the private beta of Aardvark, a security research agent, and providing free codebase scanning for major open-source projects like Next.js. To further support cyber defense, OpenAI is committing $10M in API credits to the Cybersecurity Grant Program for open-source software and critical infrastructure.
Technical Specifications and Availability
- Performance: 25% faster than previous versions due to infrastructure and inference stack improvements.
- Hardware: Co-designed, trained, and served on NVIDIA GB200 NVL72 systems.
- Availability: Available to paid ChatGPT plans via the Codex app, CLI, IDE extension, and web. API access is planned for a future release.
Performance Benchmarks
| Benchmark | GPT-5.3-Codex | GPT-5.2-Codex | GPT-5.2 |
|---|---|---|---|
| SWE-Bench Pro (Public) | 56.8% | 56.4% | 55.6% |
| Terminal-Bench 2.0 | 77.3% | 64.0% | 62.2% |
| OSWorld-Verified | 64.7% | 38.2% | 37.9% |
| GDPval (wins or ties) | 70.9% | - | 70.9% (high) |
| Cybersecurity CTF Challenges | 77.6% | 67.4% | 67.7% |
| SWE-Lancer IC Diamond | 81.4% | 76.0% | 74.6% |
Sources
- OriginalIntroducing GPT-5.3-Codex