OpenAI CEO Sam Altman has announced the official release of GPT-5.3-Codex. The new model achieves significant breakthroughs in coding performance, with scores of 57% on the SWE-Bench Pro evaluation, 76% on TerminalBench 2.0, and 64% on OSWorld.
It introduces support for real-time control and dynamic updates during task execution. Operational efficiency has also seen a substantial leap; for completing equivalent tasks, token usage is reduced by more than half compared to the 5.2-Codex version, while single-token processing speed has increased by over 25%.
The model possesses mature computer operation capabilities. Altman also stated that this is their first model to receive a "high-level" rating in the cybersecurity dimension within the safety protection framework. Currently, they are piloting a trusted access framework and investing $10 million in API credits to accelerate the development of cybersecurity defense systems.