OpenAI has unveiled GPT-5.3-Codex, its latest code-focused model, boasting a speed increase of 25% over its predecessor, GPT-5.2-Codex. It excels in software engineering tasks, achieving 56.8% accuracy on the SWE-bench Pro benchmark and a notable 77.3% on Terminal-Bench 2.0, a significant jump from 64.0%. This model also demonstrates strong performance in computer vision tasks, scoring 64.7% on OSWorld-Verified, closing in on human accuracy. The introduction of a “guidance” capability allows real-time developer interaction, enhancing context retention during code generation and debugging. Trained on NVIDIA GB200 NVL72 systems, GPT-5.3-Codex prioritizes performance and token efficiency. It is classified as “High Capability” for biosafety and cybersecurity, specifically trained to detect software vulnerabilities. This release signifies a shift toward autonomous engineering agents, allowing for streamlined multilingual workflows and reduced latency. Future projects may involve biometric systems to curb bot activity on social platforms.
Source link
