GPT-5-Codex Release and Safety Addendum

OpenAI has introduced GPT-5-Codex, a specialized version of the GPT-5 model optimized for agentic coding. This model is designed to generate code that aligns with human coding styles, adheres strictly to instructions, and iteratively tests its own output to ensure passing results.

Agentic Coding Capabilities

GPT-5-Codex is optimized for agentic workflows in Codex. To achieve this, the model was trained using reinforcement learning (RL) on real-world coding tasks across various environments. This training approach allows the model to:

Mirror Human Style: Generate code that reflects human coding styles and Pull Request (PR) preferences.
Instruction Adherence: Follow complex instructions with high precision.
Iterative Testing: Run tests iteratively until successful results are achieved.

Availability and Integration

GPT-5-Codex is available across multiple interfaces to support both local and cloud-based development workflows:

Local Access: Available via the Codex CLI and IDE extensions.
Cloud Access: Available through the the Codex web interface, GitHub, and the ChatGPT mobile app.

Safety and Mitigation Measures

OpenAI has implemented a multi-layered safety framework for GPT-5-Codex to mitigate risks associated with agentic code generation. These measures are divided into model-level and product-level mitigations:

Model-Level Mitigations

These mitigations are built directly into the model's training and behavior:

Specialized Safety Training: The model has undergone specific training to prevent the execution of harmful tasks.
Prompt Injection Defense: Training is included to mitigate the risk of prompt injections.

Product-Level Mitigations

These mitigations are implemented at the application and environment level:

Agent Sandboxing: The model operates within a sandbox to isolate the agent's actions from the host system.
Configurable Network Access: Network access for the agent can be configured to limit potential external risks.