News on February 6,OpenAI Yesterday, February 5th, a public announcement was issuedProclaimed the launch of GPT-5.3-Codex as its strongest AI ever AgentProgramming Model.
1AI ATTACHES THE FOLLOWING RESULTS OF THE RELEVANT BASELINE TESTS:
- In the SWE-bench Pro (public version) test, the model scored 56.8%, exceeding the previous generation ' s 56.4%
- In the Terminal-Bench 2.0 test, the score of the previous generation, 64.0%, jumped significantly to 773%
- Most notably, in the OSWorld-Verified computer performance baseline test, the score jumped from 38.2% to 64.7%, indicating a qualitative leap in its ability to operate in complex environments。
- Official data show that the model has not only reached new heights in many programming baseline tests, but has also improved its speed of operation by 251 TP3T compared to previous GPT-5.2-Codex。
OpenAI works to make GPT-5.3-Codex a better developer collaboration tool. The new model supports “direction and interaction” by users in the performance of tasks and does not lead to the loss of context。
In the Cordex application, the model updates high-frequency work progress, allowing developers to ask questions in real time, discuss options and revise directions. Official literature indicates that Codex is evolving from a single programming intelligence to a more universal computer collaborator, significantly lowering the threshold for building applications。
GPT-5.3-Codex is now fully accessible to ChatGPT-paying users, covering subscription schemes such as Plus, Team and Enterprise. Users can immediately access the Codex application, command line tool (CLI), IDE extension plugins and web-ends, and the API version will also be online in the near future。