OpenAI launched GPT-5.3-Codex as a code-focused agent and called it the most advanced Codex model it has released so far. OpenAI said the new version runs 25% faster than GPT-5.2-Codex and targets higher speed and precision for complex, tool-driven workflows.
Key Updates
- The Codex app now includes a “guidance” feature that lets users interact in real time while the agent executes multi-step tasks
- OpenAI trained and serves GPT-5.3-Codex on NVIDIA GB200 NVL72 systems through OpenAI-NVIDIA co-design efforts
- OpenAI labeled GPT-5.3-Codex “High Capability” for biosafety and cybersecurity in its Preparedness Framework and trained it to identify software vulnerabilities
As OpenAI expands Codex beyond a background coding assistant, GPT-5.3-Codex pushes development inside ChatGPT toward fully agent-driven workflows. The model executes complex, tool-based tasks with tighter feedback loops, letting developers guide actions in real time instead of waiting for static outputs. This direction builds on OpenAI’s earlier move to bring Codex directly into ChatGPT as a first-class coding agent, outlined when it first introduced the Codex AI coding agent in ChatGPT.
Reported Benchmark Results
| Benchmark | GPT-5.3-Codex | Previous reported score |
| SWE-bench Pro (Public) | 56.8% | – |
| Terminal-Bench 2.0 | 77.3% | 64.0% |
| OSWorld-Verified | 64.7% | 38.2% |
Coverage comparing GPT-5.3-Codex with Claude Opus 4.6 framed the OpenAI model as stronger on speed and precision, while it framed Opus 4.6 as geared toward reliability and collaboration. That comparison also described Opus 4.6 with a beta 1-million-token context window and sub-agent capabilities and it claimed GPT-5.3-Codex scored about 10% higher on Terminal-Bench 2.0.
