OpenAI has published a system card for GPT-5.3-Codex, documenting evaluation results and safety mitigations for its latest code-generation model [source]. The card details performance benchmarks, red-teaming exercises, and known limitations identified during pre-deployment testing.

According to the document, GPT-5.3-Codex achieved state-of-the-art scores on HumanEval and MBPP coding benchmarks, with pass@1 rates of 92.3% and 87.1% respectively. The model was trained on a dataset including public repositories, documentation, and synthetic code examples through early 2025.

OpenAI reports conducting adversarial testing focused on insecure code generation, prompt injection vulnerabilities in generated applications, and potential for generating malicious payloads. The card states that refusals were implemented for requests explicitly asking for exploit code, ransomware, or credential-harvesting scripts. However, the system card notes that "subtle reformulations of prohibited requests may still elicit compliance" and that "code correctness does not guarantee security."

The card discloses that GPT-5.3-Codex exhibits higher rates of generating deprecated API calls compared to GPT-4, particularly in Python and JavaScript contexts. OpenAI attributes this to training data recency challenges and states that mitigation efforts are ongoing.

Limitations documented include occasional generation of non-deterministic outputs for identical prompts, hallucination of non-existent library functions, and inconsistent handling of edge cases in mathematical operations. The card recommends that developers "validate all generated code through testing and review" and avoid deploying outputs directly to production environments.

OpenAI states the model is available through API access with usage policies prohibiting automated code deployment without human oversight. The system card follows the company's practice of publishing evaluation details for major model releases.