OpenAI Deploys GPT-5.3 Codex, Citing Self-Creation and Performance Gains

Edited by: Veronika Radoslavskaya

OpenAI has launched GPT-5.3 Codex, its most advanced agentic coding system to date. This iteration merges the coding prowess of GPT-5.2-Codex with the enhanced reasoning and professional knowledge from the GPT-5.2 base model into a unified powerhouse.

Key Enhancements

An early version of GPT-5.3 Codex assisted OpenAI's engineering team in debugging training runs, scaling deployments, and analyzing evaluation results during its own development. The model goes beyond code generation, acting as an autonomous assistant for multi-step tasks like automating routine coding chores and managing full project lifecycles.

Performance sees a major boost, with GPT-5.3 Codex running 25% faster than its predecessor. In independent tests, it outperforms Claude Opus 4.6 on Terminal-Bench 2.0, excelling in real-world terminal use, command execution, and error handling.

Benchmarks and Comparisons

On the OSWorld benchmark, which tests computer operation akin to human users, GPT-5.3 Codex delivers substantial gains over the prior Codex version—early reports place it in the mid-60s%. This shift emphasizes models that handle the entire development loop, not just code writing.

Classified as High capability in Cybersecurity under OpenAI's Preparedness Framework, the model comes with expanded support for defensive security research. Early SWE-Bench Pro scores hover around 56.8%, where Codex leads in execution speed.

Integration and Interface

GPT-5.3 Codex enables interactive steering mid-execution, letting users guide it without losing context—a fix for continuity issues in long tasks, contrasting competitors' focus on deep autonomous planning. The new Codex app for macOS serves as a command center for directing AI agents.

It's now rolled out across paid ChatGPT plans, prioritizing seamless integration into developer workflows.

Competitive Context

The release aligns with Anthropic's Claude Opus 4.6, which prioritizes deep reasoning and a 1-million-token context window. GPT-5.3 Codex expands scope to tackle nearly any professional computer task, from generating test suites to handover docs, transforming daily developer routines.

2 Views

Sources

  • WWWhat's new

  • The Week AI Broke Wall Street and the Gloves Came Off

  • OpenAI releases GPT-5.3-Codex, a coding model that helped build itself - Mashable

  • Our complete GPT 5.3 Codex review: A new era for agentic AI - eesel AI

  • OpenAI, Anthropic fight on the frontier - The Rundown AI

  • Introducing Claude Opus 4.6 - Anthropic

Did you find an error or inaccuracy?We will consider your comments as soon as possible.