OpenAI introduced GPT-5.3 Codex, a new model focused on software engineering, code reasoning, and agent-style development workflows.

Details:
- Faster and more efficient long context window
- Stronger reasoning across entire repositories
- Better refactoring, debugging, and code search
- Improved tool use for agent-style workflows

- ai agents learn ur work, use tools/computers, get better over time.

- auto-eval + optimize so managers see wtf works, ai improves itself.

- layers: chatgpt ent, agents (urs/openai/3rd party), model smarts for real tasks.

- connects to ur biz apps/systems, keeps shit governed n observable.

Key details:

- Scores ~77.6 on OpenCompass benchmarks, beating GPT-4o and Gemini 2.0 Pro despite only 9B parameters

- Supports voice chat, image/video analysis, and full-duplex live streaming

- 9B parameters (~23.4 GB), runs on RTX 4090, CPUs, and Apple M4 iPhone/iPad-class devices

Key details:
- Achieves top scores on OmniDocBench (94.62) and handles text, tables, formulas, charts, code blocks, stamps, and multi-language layouts.

- Outperforms traditional OCR like Tesseract and rivals closed systems on accuracy benchmarks.

- Small and efficient: runs on CPUs, consumer GPUs, and even low-resource devices at high speed.

@CoreAti