Long Horizon Tasks with Codex (developers.openai.com)

In a noteworthy experiment, OpenAI's GPT-5.3-Codex has demonstrated its capabilities in handling long-horizon coding tasks by successfully building a design tool from scratch over a continuous 25-hour session. Utilizing about 13 million tokens and generating approximately 30,000 lines of code, the model exhibited significant autonomy and coherence in following project specifications, running tests, and self-correcting errors. This experiment underscores a major shift in the AI/ML landscape: the transition from one-shot task execution to long-term, reliable, and complex project management. The implications for the AI/ML community are profound. With advancements like durable project memory and a structured agent loop—including planning, implementation, verification, and documentation—Codex can now tackle more substantial projects without requiring constant human oversight. This capability allows developers to delegate routine coding and verification tasks, enabling them to focus on higher-level design and decision-making. Such progress not only enhances efficiency in software development but also opens pathways for non-developers to engage in coding tasks, indicating a future where coding assistance becomes an integral part of more roles across various fields.