New Version of GPT-5 Optimized for Agentic Coding in Codex

OpenAI released an addendum to the GPT-5 system card introducing GPT-5-Codex. This version is optimized for agentic coding in Codex. Trained with reinforcement learning on real-world tasks, the model aims to produce human-like code, follow detailed instructions, and iteratively test until solutions are passed. GPT-5-Codex is available via the Codex CLI, IDE extensions, GitHub, web, and the ChatGPT mobile app. Safety evaluations show strong refusal rates across disallowed content and resilience to jailbreak attempts. Specific mitigations include synthetic data training to block malware-related tasks and defenses against prompt injection attacks. Product safeguards, such as sandboxed execution and configurable network access, provide layered protection. While considered high risk in biological/chemical domains, GPT-5-Codex is not rated as high in cybersecurity, but it includes precautionary measures for secure deployment.

OpenAI, Safety Publication – Addendum to GPT-5 system card: GPT-5-Codex – 15th September 2025