A
argbe.tech - news1min read
GPT-5.3-Codex extends Codex from coding to full computer workflows
OpenAI introduced GPT-5.3-Codex, a faster agentic coding model designed for long-running, tool-using work that spans the full software lifecycle.
OpenAI introduced GPT-5.3-Codex, positioning Codex as an agent that can run end-to-end developer and professional workflows on a computer.
- GPT-5.3-Codex combines GPT-5.2-Codex’s coding performance with GPT-5.2’s reasoning/professional knowledge, and is described as 25% faster.
- OpenAI says it sets a new industry high on SWE-Bench Pro and Terminal-Bench 2.0, and also reports strong results on OSWorld and GDPval.
- SWE-Bench Pro is framed as more contamination-resistant than SWE-bench Verified and spans four languages (Verified is Python-only).
- The Codex team reports early GPT-5.3-Codex variants helped debug training, manage deployment, and diagnose evaluations for the model itself.
- For long-horizon “build and iterate” testing, OpenAI had it autonomously evolve two web games (a racing game v2 and a diving game) over millions of tokens.